<?xml version="1.0" encoding="UTF-8"?><rss version="2.0"><channel><title>The AI News</title><description>The AI News 提供最新、最可靠的人工智能新闻：模型发布、工具评测、研究进展与行业动态。</description><link>https://theainews.cc/</link><language>zh-CN</language><item><title>2026 年 RAG 技术综述：从混合检索到 Agentic RAG 的演进与开源生态全景</title><link>https://theainews.cc/articles/2026-11-14-sota-rag-survey/</link><guid isPermaLink="true">https://theainews.cc/articles/2026-11-14-sota-rag-survey/</guid><description>梳理 RAG 从朴素管线到智能体化的四代范式演进，覆盖混合检索、上下文工程、图结构、多模态等关键技术及代表数据，并对比主流开源系统的选型要点。</description><pubDate>Sat, 14 Nov 2026 00:00:00 GMT</pubDate><category>research</category><category>RAG</category><category>检索增强生成</category><category>Agentic RAG</category><category>GraphRAG</category><category>上下文工程</category><category>向量数据库</category><category>开源</category></item><item><title>AI 破译哥德尔尘封手稿：663 条定理重见天日，选择公理证明只差一步</title><link>https://theainews.cc/articles/2026-11-13-godel-notebooks-ai/</link><guid isPermaLink="true">https://theainews.cc/articles/2026-11-13-godel-notebooks-ai/</guid><description>赫尔辛基大学团队用 Claude 解读哥德尔 1,009 页速记笔记，几天内完成人工需要数十年的工作，经 Coq 验证的定理约 400 条。</description><pubDate>Fri, 13 Nov 2026 00:00:00 GMT</pubDate><category>research</category><category>哥德尔</category><category>AI</category><category>数理逻辑</category><category>Claude</category><category>形式化验证</category><category>手稿破译</category></item><item><title>AMD Zen 6 实测：单线程比 Zen 5 快 21%，经典核与密集核首次在乱序引擎上分家</title><link>https://theainews.cc/articles/2026-11-11-amd-zen-6-family/</link><guid isPermaLink="true">https://theainews.cc/articles/2026-11-11-amd-zen-6-family/</guid><description>Chips and Cheese 实测 Zen 6 与 Zen 6c：EPYC 9686F 单线程领先 Zen 5 约 21%，两种核心首次在乱序引擎容量上分道。</description><pubDate>Wed, 11 Nov 2026 00:00:00 GMT</pubDate><category>research</category><category>AMD</category><category>Zen 6</category><category>Zen 6c</category><category>EPYC</category><category>CPU 微架构</category><category>SPEC CPU 2026</category><category>Chips and Cheese</category></item><item><title>Aqueduct：编码智能体改写程序并用 Lean 证明等价，189 个函数里 156 个拿到验证加速</title><link>https://theainews.cc/articles/2026-11-11-aqueduct/</link><guid isPermaLink="true">https://theainews.cc/articles/2026-11-11-aqueduct/</guid><description>Berkeley RDI 与 AWS 的 Aqueduct 让智能体改写函数并在 Lean 中证明等价：Verina 156/189 题拿到验证加速，几何平均 1.85×。</description><pubDate>Wed, 11 Nov 2026 00:00:00 GMT</pubDate><category>research</category><category>Aqueduct</category><category>Lean 4</category><category>形式化验证</category><category>程序优化</category><category>编码智能体</category><category>Berkeley RDI</category></item><item><title>CyberBinGym：只给两个二进制，AI 智能体恢复漏洞 88%、真正利用仅 20%</title><link>https://theainews.cc/articles/2026-11-11-cyberbingym/</link><guid isPermaLink="true">https://theainews.cc/articles/2026-11-11-cyberbingym/</guid><description>Berkeley RDI 等发布 CyberBinGym：100 个真实漏洞、无源码无初始输入，Codex 恢复 88%、利用 20%，暴露补丁到部署的空窗风险。</description><pubDate>Wed, 11 Nov 2026 00:00:00 GMT</pubDate><category>research</category><category>CyberBinGym</category><category>Berkeley RDI</category><category>二进制分析</category><category>漏洞利用</category><category>1-day</category><category>安全基准</category><category>智能体</category></item><item><title>OpenRouter 实测量化精度：fp8、fp4 与 fp16 在 240 道题上没测出差距</title><link>https://theainews.cc/articles/2026-11-11-llm-quantization/</link><guid isPermaLink="true">https://theainews.cc/articles/2026-11-11-llm-quantization/</guid><description>OpenRouter 在 Qwen3.8 27B 的 6 个端点上实测 fp16、fp8、fp4：240 道精确答案任务全部落在 98.3%–99.2%，精度标签未能预测准确率。</description><pubDate>Wed, 11 Nov 2026 00:00:00 GMT</pubDate><category>research</category><category>OpenRouter</category><category>量化</category><category>fp8</category><category>fp4</category><category>bf16</category><category>Qwen3.8</category><category>推理成本</category></item><item><title>Modern Web Guidance：Chrome 联合 Edge、Firefox 与 MDN，用 252 篇指南纠正智能体的老旧 Web 写法</title><link>https://theainews.cc/articles/2026-11-11-modern-web-guidance/</link><guid isPermaLink="true">https://theainews.cc/articles/2026-11-11-modern-web-guidance/</guid><description>Modern Web Guidance 把 252 篇现代 Web 指南接进编码智能体，官方评测通过率平均提升 39 个百分点，无障碍检查提升 35 个百分点。</description><pubDate>Wed, 11 Nov 2026 00:00:00 GMT</pubDate><category>tools</category><category>Modern Web Guidance</category><category>Google Chrome</category><category>Agent Skills</category><category>编码智能体</category><category>Web 平台</category><category>Baseline</category><category>前端</category></item><item><title>OpenRouter 实测 MCP、CLI 与 Skill 的上下文成本：同一个任务最多差 70 倍</title><link>https://theainews.cc/articles/2026-11-11-openrouter-mcp-cli-skills/</link><guid isPermaLink="true">https://theainews.cc/articles/2026-11-11-openrouter-mcp-cli-skills/</guid><description>OpenRouter 在同一个 GitHub 任务上实测三种取工具方式：CLI 1,421 token、Skill 3,315、GitHub MCP 默认 59,417，成本相差 17 倍。</description><pubDate>Wed, 11 Nov 2026 00:00:00 GMT</pubDate><category>research</category><category>OpenRouter</category><category>MCP</category><category>CLI</category><category>Agent Skills</category><category>上下文工程</category><category>Tool Search</category><category>token 成本</category></item><item><title>Vero 基准：给 43 个 Lean 仓库写实现并证明，最强组合也只全解 27 个</title><link>https://theainews.cc/articles/2026-11-11-vero/</link><guid isPermaLink="true">https://theainews.cc/articles/2026-11-11-vero/</guid><description>Berkeley RDI 等发布 Vero：43 个多模块 Lean 4 仓库、743 个 API、2,705 条规格；最强组合全解 27/43，单条规格通过率 87.3%。</description><pubDate>Wed, 11 Nov 2026 00:00:00 GMT</pubDate><category>research</category><category>Vero</category><category>Lean 4</category><category>形式化验证</category><category>代码生成</category><category>基准评测</category><category>Berkeley RDI</category></item><item><title>Anthropic 发布 Claude Fable 5.2 与 Mythos 5.2：同一底座、两套护栏，主打长时程自主任务</title><link>https://theainews.cc/articles/2026-11-09-anthropic-fable-mythos-5-2/</link><guid isPermaLink="true">https://theainews.cc/articles/2026-11-09-anthropic-fable-mythos-5-2/</guid><description>Anthropic 发布 Fable 5.2 与 Mythos 5.2：同一底座两套护栏，FrontierCode 2.0 达 49.1%，千个智能体 30 小时形式化费马大定理。</description><pubDate>Mon, 09 Nov 2026 00:00:00 GMT</pubDate><category>models</category><category>Anthropic</category><category>Claude Fable</category><category>Claude Mythos</category><category>模型发布</category><category>智能体编码</category><category>基准评测</category></item><item><title>Anthropic 拆解智能体记忆：日志、笔记、wiki 与「做梦」四层结构</title><link>https://theainews.cc/articles/2026-11-07-anthropic-agent-memory/</link><guid isPermaLink="true">https://theainews.cc/articles/2026-11-07-anthropic-agent-memory/</guid><description>Anthropic 把智能体记忆拆成四层：会话日志、记忆笔记、wiki 等表征，以及每次会话都注入的上下文。离线「做梦」负责纠错与整理。</description><pubDate>Sat, 07 Nov 2026 00:00:00 GMT</pubDate><category>research</category><category>智能体记忆</category><category>Anthropic</category><category>Claude</category><category>dreaming</category><category>CL-Bench</category><category>Claude Managed Agents</category><category>上下文工程</category></item><item><title>AMD 单季营收将首次超过英特尔：Q4 指引 170 亿美元，数据中心占比冲向八成</title><link>https://theainews.cc/articles/2026-11-05-amd-outsell-intel-q4/</link><guid isPermaLink="true">https://theainews.cc/articles/2026-11-05-amd-outsell-intel-q4/</guid><description>AMD Q3 营收 133.8 亿美元、数据中心 85.5 亿美元同比增 97%；Q4 指引约 170 亿美元，将首次超过英特尔的 150–160 亿美元。</description><pubDate>Thu, 05 Nov 2026 00:00:00 GMT</pubDate><category>industry</category><category>AMD</category><category>英特尔</category><category>数据中心</category><category>Instinct</category><category>EPYC</category><category>Helios</category><category>财报</category><category>AI 芯片</category></item><item><title>OpenAI 让模型造实体产品：Astra 不到一周做完一台设备的外壳、PCB、固件与界面</title><link>https://theainews.cc/articles/2026-11-05-astra-physical-engineering/</link><guid isPermaLink="true">https://theainews.cc/articles/2026-11-05-astra-physical-engineering/</guid><description>OpenAI 让模型参与实体产品设计：Astra 一周内做完一台可对话设备的外壳、PCB 与固件，硬件团队 Token 用量自 2 月增 23 倍，并与马自达合作优化风阻。</description><pubDate>Thu, 05 Nov 2026 00:00:00 GMT</pubDate><category>industry</category><category>OpenAI</category><category>GPT-6 Astra</category><category>Codex</category><category>CAD</category><category>工程设计</category><category>制造业</category><category>马自达</category><category>BenchCAD</category></item><item><title>币安复盘 64.3 亿美元失窃案：假面试、被投毒的 Homebrew 与被滥用的内部 AI 工具</title><link>https://theainews.cc/articles/2026-11-05-binance-fake-interview-ai-tooling-heist/</link><guid isPermaLink="true">https://theainews.cc/articles/2026-11-05-binance-fake-interview-ai-tooling-heist/</guid><description>币安披露史上最大加密资产失窃案细节：两轮假面试让工程师在公司电脑上跑起恶意代码库，攻击者借内部 AI 工具横向控制两名同事，最终触及钱包签名服务。</description><pubDate>Thu, 05 Nov 2026 00:00:00 GMT</pubDate><category>industry</category><category>币安</category><category>供应链攻击</category><category>社工攻击</category><category>AI 工具</category><category>加密货币</category><category>朝鲜黑客</category><category>Homebrew</category><category>网络安全</category></item><item><title>OpenAI 用智能体流水线 10 天扫出 948 个 Linux 提权漏洞：API 花费 121 万美元，单价 1279 美元</title><link>https://theainews.cc/articles/2026-11-05-linux-kernel-vuln-pipeline/</link><guid isPermaLink="true">https://theainews.cc/articles/2026-11-05-linux-kernel-vuln-pipeline/</guid><description>OpenAI 公开自主漏洞研究流水线：2308 条候选、1724 条动态复现、948 个可用的 Linux 本地提权 Exploit，API 成本 121 万美元，并发现 10 种新利用技巧。</description><pubDate>Thu, 05 Nov 2026 00:00:00 GMT</pubDate><category>research</category><category>OpenAI</category><category>Linux 内核</category><category>漏洞挖掘</category><category>本地提权</category><category>补丁生成</category><category>kernelCTF</category><category>网络安全</category><category>AI 安全</category></item><item><title>OpenAI 的 948 个补丁去了哪：LKML 实录、AI 审 AI 与维护者的「18 个月」</title><link>https://theainews.cc/articles/2026-11-05-lkml-openai-948-followup/</link><guid isPermaLink="true">https://theainews.cc/articles/2026-11-05-lkml-openai-948-followup/</guid><description>交叉核查 OpenAI 漏洞报告的内核落地：单账号 3 分钟连发 20 封补丁、mptcp 修复进入稳定分支、LLM 声明成惯例，以及 Sashiko 与维护者的态度。</description><pubDate>Thu, 05 Nov 2026 00:00:00 GMT</pubDate><category>industry</category><category>OpenAI</category><category>Linux 内核</category><category>LKML</category><category>Sashiko</category><category>漏洞修复</category><category>Greg KH</category><category>AI 安全</category><category>kernelCTF</category></item><item><title>vLLM 首日支持 Mistral Large 4：DCP 把单机 KV 扩到 2876 万 token，DSpark 把 TPOT 压到 3.5 ms</title><link>https://theainews.cc/articles/2026-10-29-vllm-mistral-large-4-day0/</link><guid isPermaLink="true">https://theainews.cc/articles/2026-10-29-vllm-mistral-large-4-day0/</guid><description>vLLM 与 Mistral 同日给出 1T 参数 ML4 的推理支持：DCP 让 TP8 的 KV 从 360 万扩到 2876 万 token，DSpark 每步接受 2.9 个 token、TPOT 从 8.84 ms 降到 3.50 ms。</description><pubDate>Thu, 29 Oct 2026 00:00:00 GMT</pubDate><category>tools</category><category>vLLM</category><category>Mistral Large 4</category><category>推测解码</category><category>DSpark</category><category>DCP</category><category>P/D 分离</category><category>NIXL</category><category>AMD MI350</category><category>TPU</category><category>HiSparse</category></item><item><title>DeepMind 发布 Genie 4：世界模型第一次有了声音</title><link>https://theainews.cc/articles/2026-10-28-genie-4-world-simulation/</link><guid isPermaLink="true">https://theainews.cc/articles/2026-10-28-genie-4-world-simulation/</guid><description>Google DeepMind 发布 Genie 4 研究预览：原生生成环境音与交互对话，物理与人物更真实，由 Gemini 3.8 Flash 规划多步事件。</description><pubDate>Wed, 28 Oct 2026 00:00:00 GMT</pubDate><category>models</category><category>Genie 4</category><category>Google DeepMind</category><category>世界模型</category><category>Gemini</category><category>Project Genie</category><category>SynthID</category><category>具身智能</category></item><item><title>智能体原生操作系统调研：从沙箱到内核，操作系统正在为智能体重做一遍</title><link>https://theainews.cc/articles/2026-10-27-agent-native-os-survey/</link><guid isPermaLink="true">https://theainews.cc/articles/2026-10-27-agent-native-os-survey/</guid><description>学术界在做 AIOS、AgentKernel 与智能体原生文件系统，产业界把隔离边界推向 microVM 与硬件强制。本文调研智能体时代操作系统、容器与沙箱的演进方向与设计原则。</description><pubDate>Tue, 27 Oct 2026 00:00:00 GMT</pubDate><category>research</category><category>智能体</category><category>操作系统</category><category>沙箱</category><category>容器</category><category>microVM</category><category>Firecracker</category><category>gVisor</category><category>智能体安全</category><category>调研</category></item><item><title>编程智能体接管已登录 Chrome：六种方案的原理、配置与取舍</title><link>https://theainews.cc/articles/2026-10-26-coding-agent-logged-in-chrome/</link><guid isPermaLink="true">https://theainews.cc/articles/2026-10-26-coding-agent-logged-in-chrome/</guid><description>Claude Code、Codex、Gemini CLI、Cursor 如何使用你日常已登录的 Chrome？对比 Claude in Chrome、Playwright MCP、Chrome DevTools MCP、Browser Harness 等六种方案的配置与安全边界。</description><pubDate>Mon, 26 Oct 2026 00:00:00 GMT</pubDate><category>tutorial</category><category>编程智能体</category><category>MCP</category><category>Chrome</category><category>Claude Code</category><category>Gemini CLI</category><category>Codex</category><category>Cursor</category><category>浏览器自动化</category></item><item><title>ECC：把编码智能体装进一套工程操作系统</title><link>https://theainews.cc/articles/2026-10-26-ecc-agent-harness-os/</link><guid isPermaLink="true">https://theainews.cc/articles/2026-10-26-ecc-agent-harness-os/</guid><description>开源项目 ECC 用 71 个子智能体、311 个技能与强制流程，让 Claude Code、Codex 等编码智能体先计划、再测试、后复查，并跨工具共享记忆。</description><pubDate>Mon, 26 Oct 2026 00:00:00 GMT</pubDate><category>tools</category><category>ECC</category><category>智能体</category><category>Claude Code</category><category>Codex</category><category>TDD</category><category>开发工具</category><category>开源</category></item><item><title>Impeccable：把设计总监装进编程智能体，用 24 条命令和 62 条规则治「AI 味」</title><link>https://theainews.cc/articles/2026-10-26-impeccable-design-skill/</link><guid isPermaLink="true">https://theainews.cc/articles/2026-10-26-impeccable-design-skill/</guid><description>8.8 万 star 的开源设计技能 Impeccable：24 条命令、62 条确定性检测规则、浏览器里实时改稿，让编程智能体摆脱同一套 AI 设计模板。</description><pubDate>Mon, 26 Oct 2026 00:00:00 GMT</pubDate><category>tools</category><category>Impeccable</category><category>设计技能</category><category>Claude Code</category><category>Cursor</category><category>Codex</category><category>AI slop</category><category>检测规则</category><category>Rust</category><category>开源</category></item><item><title>全面介绍 Omarchy：DHH 打造的「美丽、有趣、智能体原生」Linux 发行版</title><link>https://theainews.cc/articles/2026-10-26-omarchy-intro/</link><guid isPermaLink="true">https://theainews.cc/articles/2026-10-26-omarchy-intro/</guid><description>基于 Arch 与 Hyprland 的键盘优先桌面，开箱自带 20 余家 AI 智能体 CLI 与主题系统，一年售出 130 万 ISO，基金会募资超 2300 万美元，本文从架构、体验、AI 集成、安全争议到社区生态全面解读。</description><pubDate>Mon, 26 Oct 2026 00:00:00 GMT</pubDate><category>tools</category><category>Omarchy</category><category>Linux</category><category>Arch</category><category>Hyprland</category><category>Quickshell</category><category>DHH</category><category>AI智能体</category><category>开源</category><category>桌面操作系统</category></item><item><title>OpenSandbox 加入 AAIF：阿里开源的智能体沙箱运行时走向中立治理</title><link>https://theainews.cc/articles/2026-10-26-opensandbox-aaif-sandbox-runtime/</link><guid isPermaLink="true">https://theainews.cc/articles/2026-10-26-opensandbox-aaif-sandbox-runtime/</guid><description>阿里巴巴把开源沙箱平台 OpenSandbox 捐给 Linux 基金会 AAIF；统一 API 覆盖 Docker 与 Kubernetes，Firecracker 沙箱创建 P50 64 毫秒。</description><pubDate>Mon, 26 Oct 2026 00:00:00 GMT</pubDate><category>tools</category><category>OpenSandbox</category><category>AAIF</category><category>沙箱</category><category>智能体基础设施</category><category>Firecracker</category><category>Kubernetes</category><category>开源</category><category>Linux 基金会</category></item><item><title>NVIDIA 开源 SkillSpector：安装智能体技能之前，先做一次安全体检</title><link>https://theainews.cc/articles/2026-10-26-skillspector-agent-skill-scanner/</link><guid isPermaLink="true">https://theainews.cc/articles/2026-10-26-skillspector-agent-skill-scanner/</guid><description>NVIDIA 开源技能安全扫描器 SkillSpector：71 条检测规则、静态分析加可选 LLM 语义复核、0-100 风险分；论文测得 26.1% 的技能至少含一个漏洞。</description><pubDate>Mon, 26 Oct 2026 00:00:00 GMT</pubDate><category>tools</category><category>SkillSpector</category><category>NVIDIA</category><category>智能体技能</category><category>安全扫描</category><category>提示词注入</category><category>MCP</category><category>供应链安全</category><category>开源</category></item><item><title>JetBrains 开源 ThinkRail：给极简的 pi 智能体套上一个真正的 IDE</title><link>https://theainews.cc/articles/2026-10-26-thinkrail-jetbrains-agentic-ide/</link><guid isPermaLink="true">https://theainews.cc/articles/2026-10-26-thinkrail-jetbrains-agentic-ide/</guid><description>JetBrains 开源智能体 IDE ThinkRail：每个工作区是一个 git worktree，改动按轮次出收据，项目意图写进规格图，底层跑的是 pi。</description><pubDate>Mon, 26 Oct 2026 00:00:00 GMT</pubDate><category>tools</category><category>ThinkRail</category><category>JetBrains</category><category>pi</category><category>智能体 IDE</category><category>git worktree</category><category>规格驱动开发</category><category>MCP</category><category>开源</category></item><item><title>SGLang 对标 DeepInfer：同硬件跑对方的基准，prefill 最高 1.04 倍、64 卡 decode 单步短 7.4%</title><link>https://theainews.cc/articles/2026-10-24-sglang-deepinfer-benchmark/</link><guid isPermaLink="true">https://theainews.cc/articles/2026-10-24-sglang-deepinfer-benchmark/</guid><description>SGLang 团队用 DeepInfer 自己的基准在相同硬件上实测：prefill 吞吐最高 1.04 倍，64 卡 decode 单步时间短 1.1%–7.4%，并补齐 DeepInfer 未开源的 KV 缓存、PD 分离与副本负载均衡。</description><pubDate>Sat, 24 Oct 2026 00:00:00 GMT</pubDate><category>tools</category><category>SGLang</category><category>DeepInfer</category><category>DeepSeek</category><category>V4.1-Flash</category><category>推理优化</category><category>MegaMoE</category><category>投机解码</category><category>KV 缓存</category></item><item><title>vLLM 用 P/D 分离与 Wide-EP 跑满一个 GB300 NVL72 机柜：单卡 954K tokens/s，decode 差 DeepInfer 6% 以内</title><link>https://theainews.cc/articles/2026-10-23-vllm-deepseek-v41-flash-nvl72/</link><guid isPermaLink="true">https://theainews.cc/articles/2026-10-23-vllm-deepseek-v41-flash-nvl72/</guid><description>vLLM 在单个 GB300 NVL72 机柜上服务 V4.1-Flash 达单卡 954K tokens/s、每 GPU 小时 38.96 美元；decode 吞吐与 DeepInfer 差 6% 以内，prefill 三项优化再提 43.9%。</description><pubDate>Fri, 23 Oct 2026 00:00:00 GMT</pubDate><category>tools</category><category>vLLM</category><category>DeepSeek</category><category>V4.1-Flash</category><category>DeepInfer</category><category>GB300 NVL72</category><category>P/D 分离</category><category>Wide-EP</category><category>EPLB</category><category>推理优化</category></item><item><title>Browser Use 拆解浏览器智能体：三层结构、真实用量与五类失效模式</title><link>https://theainews.cc/articles/2026-10-21-browser-agents-explainer/</link><guid isPermaLink="true">https://theainews.cc/articles/2026-10-21-browser-agents-explainer/</guid><description>Browser Use 发文拆解浏览器智能体：模型决策、执行框架跑循环、浏览器持有会话；并给出「API 优先、脚本其次、智能体最后」的选型顺序与五类失效模式。</description><pubDate>Wed, 21 Oct 2026 00:00:00 GMT</pubDate><category>tutorial</category><category>浏览器智能体</category><category>Browser Use</category><category>Computer Use</category><category>Playwright</category><category>MCP</category><category>提示注入</category><category>智能体工程</category></item><item><title>蚂蚁集团与 vLLM 社区开源 Humming：统一量化推理内核，端到端吞吐最高提升 43%</title><link>https://theainews.cc/articles/2026-10-21-humming-quantized-inference-kernels/</link><guid isPermaLink="true">https://theainews.cc/articles/2026-10-21-humming-quantized-inference-kernels/</guid><description>蚂蚁集团与 vLLM 社区开源量化推理内核库 Humming：统一量化描述 + JIT 编译覆盖 INT1 到 FP8，8 卡 H20 上端到端吞吐最高提升 43%。</description><pubDate>Wed, 21 Oct 2026 00:00:00 GMT</pubDate><category>tools</category><category>Humming</category><category>vLLM</category><category>蚂蚁集团</category><category>量化推理</category><category>GEMM</category><category>MoE</category><category>CUDA</category><category>JIT 编译</category><category>MXFP4</category><category>Blackwell</category></item><item><title>Apate.ai 用 AI 智能体反钓骗子：累计耗掉 17 年诈骗时间，携 815 万美元融资进军美国</title><link>https://theainews.cc/articles/2026-10-20-apate-ai-scam-baiting-agents/</link><guid isPermaLink="true">https://theainews.cc/articles/2026-10-20-apate-ai-scam-baiting-agents/</guid><description>澳洲初创 Apate.ai 用数千个 AI 智能体冒充受害者反钓骗子，累计耗掉 17 年诈骗时间；8 月获 815 万美元种子轮后进军美国，专家提醒「耗时间」不等于损失下降。</description><pubDate>Tue, 20 Oct 2026 00:00:00 GMT</pubDate><category>industry</category><category>Apate.ai</category><category>反诈</category><category>AI 智能体</category><category>欺诈情报</category><category>网络安全</category><category>诈骗防御</category><category>TechCrunch Disrupt</category></item><item><title>美陆军网络司令部被曝研发 ECHOTRIBBLE：让 AI 智能体自主导航并「中和」目标</title><link>https://theainews.cc/articles/2026-10-20-army-cyber-command-echotribble/</link><guid isPermaLink="true">https://theainews.cc/articles/2026-10-20-army-cyber-command-echotribble/</guid><description>The Intercept 获得的文件显示，美陆军网络司令部正研发代号 ECHOTRIBBLE 的 AI 智能体平台，要求以最小人工监督自主「中和」目标，多位专家警告附带损伤风险。</description><pubDate>Tue, 20 Oct 2026 00:00:00 GMT</pubDate><category>industry</category><category>AI 智能体</category><category>网络安全</category><category>军事 AI</category><category>自主武器</category><category>ARCYBER</category><category>ECHOTRIBBLE</category><category>AI 安全</category></item><item><title>DeepSeek 开源生产推理框架 DeepInfer：单集群日处理 155 万亿 token，NVIDIA 与昇腾同栈</title><link>https://theainews.cc/articles/2026-10-20-deepinfer-deepseek-inference-framework/</link><guid isPermaLink="true">https://theainews.cc/articles/2026-10-20-deepinfer-deepseek-inference-framework/</guid><description>DeepSeek 开源驱动 V4.1-Flash 线上服务的推理框架 DeepInfer，日处理 155 万亿 token；Wide-EP 解码、SWA 有界重放与零开销调度齐备，同时支持 NVIDIA 与昇腾。</description><pubDate>Tue, 20 Oct 2026 00:00:00 GMT</pubDate><category>tools</category><category>DeepSeek</category><category>DeepInfer</category><category>V4.1-Flash</category><category>推理框架</category><category>开源</category><category>Wide-EP</category><category>EPLB</category><category>SWA 有界重放</category><category>投机解码</category><category>KV 缓存</category></item><item><title>Agent Client Protocol 一年：42 款编程智能体接入，编辑器与智能体正在解耦</title><link>https://theainews.cc/articles/2026-10-18-agent-client-protocol-explainer/</link><guid isPermaLink="true">https://theainews.cc/articles/2026-10-18-agent-client-protocol-explainer/</guid><description>ACP 让编程智能体和编辑器通过一套 JSON-RPC 协议互相适配。发布一年后注册表已有 42 款智能体，v1 与 SDK 稳定，远程接入仍在进行中。</description><pubDate>Sun, 18 Oct 2026 00:00:00 GMT</pubDate><category>industry</category><category>Agent Client Protocol</category><category>ACP</category><category>AI 编程智能体</category><category>Zed</category><category>JetBrains</category><category>IDE</category><category>互操作标准</category><category>LSP</category></item><item><title>AgentRQ 发布 Pano：给所有工作区的智能体一块共享的 Markdown 留言板</title><link>https://theainews.cc/articles/2026-10-18-agentrq-pano-shared-board/</link><guid isPermaLink="true">https://theainews.cc/articles/2026-10-18-agentrq-pano-shared-board/</guid><description>AgentRQ v0.10.0 推出 Pano：账户内所有工作区共享一块 Markdown 文档板，智能体可发帖、评论、互相交接，清单项完成还会自动打勾。</description><pubDate>Sun, 18 Oct 2026 00:00:00 GMT</pubDate><category>tools</category><category>AgentRQ</category><category>多智能体协作</category><category>MCP</category><category>工作区</category><category>人在回路</category><category>Markdown</category><category>任务编排</category></item><item><title>让定时智能体别翻车：Anthropic 拆解智能体自动化的六个部件与六条规则</title><link>https://theainews.cc/articles/2026-10-17-claude-managed-agent-automations/</link><guid isPermaLink="true">https://theainews.cc/articles/2026-10-17-claude-managed-agent-automations/</guid><description>Anthropic 用 Claude Managed Agents 搭了一个每日简报智能体，并总结出按书签读书、读取失败要如实报告、发布前复核、预算封顶等六条规则。</description><pubDate>Sat, 17 Oct 2026 00:00:00 GMT</pubDate><category>tutorial</category><category>Claude Managed Agents</category><category>智能体自动化</category><category>定时任务</category><category>Agent 记忆</category><category>护栏</category><category>Anthropic</category><category>MCP</category><category>提示注入</category></item><item><title>SGLang 为 Muse Spark Open 做 Day-0：序列并行 + NVFP4 MegaMoE，长提示首字延迟降 28–42%</title><link>https://theainews.cc/articles/2026-10-17-muse-spark-open-sglang-day0/</link><guid isPermaLink="true">https://theainews.cc/articles/2026-10-17-muse-spark-open-sglang-day0/</guid><description>SGLang 与 Miles 同日支持 Muse Spark Open：序列并行 + NVFP4 MegaMoE 让长提示 p90 首字延迟降 28.3–41.9%，12 步内核优化把解码吞吐提到 3 倍。</description><pubDate>Sat, 17 Oct 2026 00:00:00 GMT</pubDate><category>tools</category><category>SGLang</category><category>Miles</category><category>Muse Spark Open</category><category>序列并行</category><category>NVFP4</category><category>MegaMoE</category><category>投机解码</category><category>强化学习</category></item><item><title>Reflection 开源 Beam：501B 总参 / 23B 激活，推理算力省 3–4 倍</title><link>https://theainews.cc/articles/2026-10-17-reflection-beam-open-weights/</link><guid isPermaLink="true">https://theainews.cc/articles/2026-10-17-reflection-beam-open-weights/</guid><description>Reflection 开源首个模型 Beam：501B 总参 / 23B 激活 MoE，1M 上下文、Apache 2.0，编码与智能体能力对标 GLM 5.2，推理算力少 3–4 倍。</description><pubDate>Sat, 17 Oct 2026 00:00:00 GMT</pubDate><category>models</category><category>Reflection</category><category>Beam</category><category>开源权重</category><category>MoE</category><category>高算力强化学习</category><category>Apache 2.0</category><category>智能体</category><category>NVFP4</category></item><item><title>YOLO 模式全解：12 款编程智能体的免确认开关怎么开</title><link>https://theainews.cc/articles/2026-10-17-yolo-mode-coding-agents/</link><guid isPermaLink="true">https://theainews.cc/articles/2026-10-17-yolo-mode-coding-agents/</guid><description>YOLO 模式让编程智能体不再逐步征求同意。梳理 12 款主流智能体的开启方式、折中档位，以及它真正关掉的东西。</description><pubDate>Sat, 17 Oct 2026 00:00:00 GMT</pubDate><category>tutorial</category><category>YOLO 模式</category><category>编程智能体</category><category>Claude Code</category><category>Codex</category><category>Gemini CLI</category><category>权限审批</category><category>沙箱</category></item><item><title>Claude Code 团队如何评审代码：把阅读交给机器，人类只做决定</title><link>https://theainews.cc/articles/2026-10-16-claude-code-review/</link><guid isPermaLink="true">https://theainews.cc/articles/2026-10-16-claude-code-review/</guid><description>Anthropic 复盘 Claude Code 的代码评审流程：六层自动检查加按风险分级的审批，6 月至 9 月合并 PR 翻倍，回滚比例降逾 40%。</description><pubDate>Fri, 16 Oct 2026 00:00:00 GMT</pubDate><category>tools</category><category>Claude Code</category><category>Anthropic</category><category>代码评审</category><category>AI 编程</category><category>CI</category><category>工程实践</category><category>CLAUDE.md</category><category>Agent</category></item><item><title>Kimi K3 跑上 AMD MI355X：单卡吞吐超 NVIDIA HGX B300 24.4%</title><link>https://theainews.cc/articles/2026-10-16-kimi-k3-mi355x-amd/</link><guid isPermaLink="true">https://theainews.cc/articles/2026-10-16-kimi-k3-mi355x-amd/</guid><description>AMD 与 vLLM 社区复盘 Kimi K3 在 MI355X 上的优化：AgentX 基准单卡吞吐超 NVIDIA HGX B300 24.4%，每美元 token 多 87.4%。</description><pubDate>Fri, 16 Oct 2026 00:00:00 GMT</pubDate><category>industry</category><category>vLLM</category><category>Kimi K3</category><category>AMD MI355X</category><category>ROCm</category><category>推理优化</category><category>投机解码</category><category>MoE</category><category>AgentX</category></item><item><title>Meta 开源 Muse Spark Open：816B 参数、100 万上下文，2 台 Mac Studio 就能跑</title><link>https://theainews.cc/articles/2026-10-16-meta-muse-spark-open/</link><guid isPermaLink="true">https://theainews.cc/articles/2026-10-16-meta-muse-spark-open/</guid><description>Meta 开源 Muse Spark Open：816B 总参 / 42B 激活的 MoE，1M 上下文与文本、图像、视频、音频输入，NVFP4 版可在 2 台 M5 Ultra Mac Studio 上本地运行。</description><pubDate>Fri, 16 Oct 2026 00:00:00 GMT</pubDate><category>models</category><category>Meta</category><category>Muse Spark Open</category><category>开源权重</category><category>MoE</category><category>多模态</category><category>NVFP4</category><category>MLX</category><category>投机解码</category></item><item><title>上下文工程实战：LangChain 拆解 Deep Agents 如何「记住」而不塞满窗口</title><link>https://theainews.cc/articles/2026-10-15-context-engineering-deep-agents/</link><guid isPermaLink="true">https://theainews.cc/articles/2026-10-15-context-engineering-deep-agents/</guid><description>LangChain 撰文拆解 Deep Agents 的上下文工程：卸载、摘要、技能、子智能体与长期记忆五种中间件，如何在不动原始数据的前提下把上下文移出活动窗口。</description><pubDate>Thu, 15 Oct 2026 00:00:00 GMT</pubDate><category>tools</category><category>LangChain</category><category>Deep Agents</category><category>上下文工程</category><category>Agent Harness</category><category>中间件</category><category>长期记忆</category></item><item><title>OpenCode 2 正式发布：API 全部重做、Code Mode 上线，终端与桌面共用一个后台服务</title><link>https://theainews.cc/articles/2026-10-15-opencode-2/</link><guid isPermaLink="true">https://theainews.cc/articles/2026-10-15-opencode-2/</guid><description>OpenCode 2 正式发布：终端、桌面、网页共用同一后台服务，新增 Code Mode 把 MCP 与 OpenAPI 编译成 JavaScript 函数，插件 API 与服务端 API 均为不兼容重写。</description><pubDate>Thu, 15 Oct 2026 00:00:00 GMT</pubDate><category>tools</category><category>OpenCode</category><category>AI 编程智能体</category><category>Code Mode</category><category>MCP</category><category>插件系统</category><category>迁移</category></item><item><title>Gemini 4 Argon 结束定向预览：API 转 Stable，限时价格 2/10 美元只有七周</title><link>https://theainews.cc/articles/2026-10-14-gemini-4-argon-ga/</link><guid isPermaLink="true">https://theainews.cc/articles/2026-10-14-gemini-4-argon-ga/</guid><description>Gemini 4 Argon 正式上线：Gemini API 转 Stable、进入 Gemini 应用与 AI 搜索模式，限时价每百万输入 2 美元、输出 10 美元，12 月 1 日起翻倍为 4/20 美元。</description><pubDate>Wed, 14 Oct 2026 00:00:00 GMT</pubDate><category>models</category><category>Google DeepMind</category><category>Gemini 4 Argon</category><category>前沿模型</category><category>API 定价</category><category>Google Antigravity</category><category>Token 效率</category><category>Artificial Analysis</category></item><item><title>智谱发布 GLM-Legal：Harvey LAB 全通过率 19.2%，超过同门基座模型 12.5 个百分点</title><link>https://theainews.cc/articles/2026-10-12-glm-legal/</link><guid isPermaLink="true">https://theainews.cc/articles/2026-10-12-glm-legal/</guid><description>智谱发布面向法律工作的 GLM-Legal：基于 GLM-5.3-Flash，在 Harvey LAB 取得 19.2% 全通过率、LegalBench 86.2%，以远小于闭源前沿模型的参数规模打平，API 定价每百万 token 输入 0.075 美元。</description><pubDate>Mon, 12 Oct 2026 00:00:00 GMT</pubDate><category>models</category><category>智谱</category><category>GLM-Legal</category><category>GLM-5.3-Flash</category><category>法律 AI</category><category>垂直模型</category><category>Agent Harness</category><category>Harvey LAB</category><category>LegalBench</category><category>AI Agent</category></item><item><title>Meta 发布首个前沿模型 Muse Wonder：用 Active Memory 与 Active Focus 做个人智能体</title><link>https://theainews.cc/articles/2026-10-12-muse-wonder/</link><guid isPermaLink="true">https://theainews.cc/articles/2026-10-12-muse-wonder/</guid><description>Meta 推出 Muse Wonder 与 Wonder Full：100 万 token 上下文的首个前沿模型，API 定价每百万输入 4 美元、输出 20 美元。</description><pubDate>Mon, 12 Oct 2026 00:00:00 GMT</pubDate><category>models</category><category>Meta</category><category>Muse Wonder</category><category>前沿模型</category><category>AI Agent</category><category>Active Memory</category><category>模型安全</category><category>Prometheus</category></item><item><title>vLLM 优化 DeepSeek-V4.1-Flash：智能体吞吐 5.3 倍，TTFT 降近七成</title><link>https://theainews.cc/articles/2026-10-11-vllm-deepseek-v41-flash/</link><guid isPermaLink="true">https://theainews.cc/articles/2026-10-11-vllm-deepseek-v41-flash/</guid><description>vLLM 团队复盘 DeepSeek-V4.1-Flash 发布后三周的推理优化：SWA 有界重放加 CUDA 图把 prefill TTFT 降三成，内核全链路融合让 AgentX 吞吐提升 5.3 倍。</description><pubDate>Sun, 11 Oct 2026 00:00:00 GMT</pubDate><category>tools</category><category>vLLM</category><category>DeepSeek</category><category>V4.1-Flash</category><category>推理优化</category><category>SWA</category><category>CUDA Graph</category><category>内核融合</category><category>AgentX</category></item><item><title>Microsoft 开源 MXC：给不可信代码与智能体输出用的跨平台沙箱，发布 1.0</title><link>https://theainews.cc/articles/2026-10-09-microsoft-mxc/</link><guid isPermaLink="true">https://theainews.cc/articles/2026-10-09-microsoft-mxc/</guid><description>Microsoft 开源 MXC：跨 Windows、Linux、macOS 的沙箱 SDK，用统一策略模型把隔离从进程沙箱做到微 VM，已发布 1.0。</description><pubDate>Fri, 09 Oct 2026 00:00:00 GMT</pubDate><category>tools</category><category>Microsoft</category><category>MXC</category><category>沙箱</category><category>代码执行</category><category>AI Agent</category><category>安全隔离</category><category>Rust</category><category>SDK</category><category>开源项目</category></item><item><title>Anthropic 发布 Claude Haiku 5.5：最便宜最快的小模型，平均便宜约 75%</title><link>https://theainews.cc/articles/2026-10-08-claude-haiku-5-5/</link><guid isPermaLink="true">https://theainews.cc/articles/2026-10-08-claude-haiku-5-5/</guid><description>Anthropic 发布 Claude Haiku 5.5：Haiku 档首个可调 effort 小模型，每百万输入 0.1 美元、输出 0.5 美元起。</description><pubDate>Thu, 08 Oct 2026 00:00:00 GMT</pubDate><category>models</category><category>Anthropic</category><category>Claude Haiku 5.5</category><category>小模型</category><category>模型定价</category><category>可调 effort</category><category>AI Agent</category><category>AI 安全</category></item><item><title>Docker 开源 docker-agent：用一份 YAML 声明多智能体，并把它推到 OCI 仓库共享</title><link>https://theainews.cc/articles/2026-10-08-docker-agent/</link><guid isPermaLink="true">https://theainews.cc/articles/2026-10-08-docker-agent/</guid><description>Docker 开源的 docker-agent 用 YAML 声明模型、工具与子智能体来跑多智能体协作，还能把智能体打包推到 OCI 仓库共享。</description><pubDate>Thu, 08 Oct 2026 00:00:00 GMT</pubDate><category>tools</category><category>Docker</category><category>docker-agent</category><category>AI Agent</category><category>MCP</category><category>多智能体</category><category>RAG</category><category>YAML</category><category>OCI</category><category>开源项目</category></item><item><title>OpenAI 把 GPT-6 和 Intelligent UI 推向所有人：免费档明日起可用</title><link>https://theainews.cc/articles/2026-10-08-gpt-6-intelligent-ui/</link><guid isPermaLink="true">https://theainews.cc/articles/2026-10-08-gpt-6-intelligent-ui/</guid><description>OpenAI 把 GPT-6 与 Intelligent UI 推向全体 ChatGPT 用户：免费档明天起开放，付费档由 GPT-6 Sol 驱动。</description><pubDate>Thu, 08 Oct 2026 00:00:00 GMT</pubDate><category>models</category><category>OpenAI</category><category>GPT-6</category><category>Intelligent UI</category><category>ChatGPT</category><category>GPT-6 Sol</category><category>GPT-6 Luna</category><category>生成式 UI</category><category>AI 安全</category></item><item><title>Mistral 发布 Large 4 预览：1 万亿参数 MoE，主打开源权重与欧洲主权 AI</title><link>https://theainews.cc/articles/2026-10-07-mistral-large-4/</link><guid isPermaLink="true">https://theainews.cc/articles/2026-10-07-mistral-large-4/</guid><description>Mistral 公开预览 1 万亿参数、490 亿激活的原生多模态 MoE 模型 Large 4，权重月底开放，并以网络安全、金融、法律等企业场景和欧洲主权 AI 为卖点。</description><pubDate>Wed, 07 Oct 2026 00:00:00 GMT</pubDate><category>models</category><category>Mistral</category><category>Mistral Large 4</category><category>MoE</category><category>开源权重</category><category>AI 主权</category><category>网络安全</category><category>多模态</category><category>强化学习</category></item><item><title>OpenAI 上线 Decisions API：用 gpt-6-luna 换回「一个概率」而不是一段话</title><link>https://theainews.cc/articles/2026-10-07-openai-decisions-api/</link><guid isPermaLink="true">https://theainews.cc/articles/2026-10-07-openai-decisions-api/</guid><description>OpenAI 公开测试新版 Decisions API：开发者用 predicate、choice、score 三种问题类型，让 gpt-6-luna 对文本或图像返回概率化答案，官方称比 Responses API 快约 10 倍。</description><pubDate>Wed, 07 Oct 2026 00:00:00 GMT</pubDate><category>tools</category><category>OpenAI</category><category>Decisions API</category><category>gpt-6-luna</category><category>API</category><category>结构化输出</category><category>内容分类</category><category>路由</category><category>数据驻留</category></item><item><title>什么是 Codemode：Armin Ronacher 谈智能体外壳侧的编排层</title><link>https://theainews.cc/articles/2026-10-07-what-is-codemode/</link><guid isPermaLink="true">https://theainews.cc/articles/2026-10-07-what-is-codemode/</guid><description>Armin Ronacher 撰文解释 Codemode：跑在智能体外壳沙箱里、用 JavaScript 组合工具调用的编排层，以及它为何是 bash 之外的必要补充。</description><pubDate>Wed, 07 Oct 2026 00:00:00 GMT</pubDate><category>tools</category><category>Codemode</category><category>Pi</category><category>MCP</category><category>agent harness</category><category>Armin Ronacher</category><category>QuickJS</category></item><item><title>AMD 计划 2027 年大幅提升芯片供应，Lisa Su 赴台北锁定产能</title><link>https://theainews.cc/articles/2026-10-06-amd-2027-supply-increase-lisa-su-taipei/</link><guid isPermaLink="true">https://theainews.cc/articles/2026-10-06-amd-2027-supply-increase-lisa-su-taipei/</guid><description>Lisa Su 在台北称 AMD 2026 年已在扩产、2027 年将大幅提升供应；此行会晤富士康、将见台积电，再赴韩国谈内存，公司已按 3–5 年提前规划晶圆产能。</description><pubDate>Tue, 06 Oct 2026 00:00:00 GMT</pubDate><category>industry</category><category>AMD</category><category>Lisa Su</category><category>台积电</category><category>富士康</category><category>HBM</category><category>AI 芯片供应</category><category>三星</category><category>SK 海力士</category></item><item><title>NVIDIA Olympus 实测：用 3.3 GHz 和「空间多线程」冲服务器单线程性能上限</title><link>https://theainews.cc/articles/2026-10-06-nvidia-olympus-core/</link><guid isPermaLink="true">https://theainews.cc/articles/2026-10-06-nvidia-olympus-core/</guid><description>Chips and Cheese 实测 NVIDIA Vera 的 Olympus：10 宽乱序、3.3 GHz，单线程逼近桌面旗舰，但「空间多线程」在部分浮点负载出现负增益。</description><pubDate>Tue, 06 Oct 2026 00:00:00 GMT</pubDate><category>research</category><category>NVIDIA</category><category>Olympus</category><category>Vera CPU</category><category>CPU 微架构</category><category>SMT</category><category>SPEC CPU</category><category>Arm X925</category><category>Zen 5</category><category>智能体 AI</category><category>SOCAMM2</category></item><item><title>SemiAnalysis 实测各家订阅：同档 Claude 的用量价值约为 OpenAI 的 5 倍</title><link>https://theainews.cc/articles/2026-10-06-semianalysis-ai-subscription-value/</link><guid isPermaLink="true">https://theainews.cc/articles/2026-10-06-semianalysis-ai-subscription-value/</guid><description>SemiAnalysis 用限额实测量化各家 AI 订阅：中端模型档 Claude 的 API 等价价值约为 OpenAI 5 倍，而 OpenAI 上周把 200 美元档额度砍半、新增 500 美元档。</description><pubDate>Tue, 06 Oct 2026 00:00:00 GMT</pubDate><category>industry</category><category>SemiAnalysis</category><category>Anthropic</category><category>OpenAI</category><category>订阅经济</category><category>Claude</category><category>GPT-6</category><category>tokenomics</category></item><item><title>Wolfram 长文：AI 会解题之后，纯数学研究靠什么继续</title><link>https://theainews.cc/articles/2026-10-06-wolfram-future-pure-math-ai/</link><guid isPermaLink="true">https://theainews.cc/articles/2026-10-06-wolfram-future-pure-math-ai/</guid><description>Wolfram 发文回应「数学研究可以全交给 AI」：AI 擅长检索与连接已有成果，但纯数学的核心是提出问题、设定目标与构建概念，这仍然只能来自人。</description><pubDate>Tue, 06 Oct 2026 00:00:00 GMT</pubDate><category>research</category><category>Stephen Wolfram</category><category>纯数学</category><category>AI 与数学</category><category>自动定理证明</category><category>形式化</category><category>计算不可约性</category><category>Wolfram Language</category></item><item><title>Muse 并无新技术：Meta 赢在产品设计与广告模式</title><link>https://theainews.cc/articles/2026-10-05-meta-muse-consumer-ai-design/</link><guid isPermaLink="true">https://theainews.cc/articles/2026-10-05-meta-muse-consumer-ai-design/</guid><description>产品设计师 Mete Polat 拆解 Muse：核心能力一年前就已存在，Meta 靠统一的心智模型、广告供血的免费层和拟人化设计，第一次让消费级智能体跑通。</description><pubDate>Mon, 05 Oct 2026 00:00:00 GMT</pubDate><category>industry</category><category>Meta</category><category>Muse</category><category>消费级 AI</category><category>产品设计</category><category>智能体</category><category>广告模式</category><category>OpenAI dots</category><category>Google</category></item><item><title>Red Hat 实测 9 种 AI 护栏：决策模型没能兑现「更快更便宜」</title><link>https://theainews.cc/articles/2026-10-05-redhat-decision-models-guardrail-benchmark/</link><guid isPermaLink="true">https://theainews.cc/articles/2026-10-05-redhat-decision-models-guardrail-benchmark/</guid><description>Red Hat AI Safety 团队在 NeMo Guardrails 上对比 4 类共 9 套护栏方案：Jev 等决策模型没有稳定胜过传统分类器或 LLM 裁判，200M 参数的预训练小模型依然最快最准。</description><pubDate>Mon, 05 Oct 2026 00:00:00 GMT</pubDate><category>research</category><category>AI 护栏</category><category>决策模型</category><category>Jev</category><category>TypeSafe</category><category>NeMo Guardrails</category><category>Red Hat</category><category>LLM-as-a-judge</category><category>提示注入</category></item><item><title>Meta 开源 Muse Gadgets：把个人 AI 智能体接进你自己的硬件</title><link>https://theainews.cc/articles/2026-10-03-muse-gadgets/</link><guid isPermaLink="true">https://theainews.cc/articles/2026-10-03-muse-gadgets/</guid><description>Meta 开源 Muse Gadgets 硬件 SDK：闲置的 ESP32 开发板或树莓派刷上固件，就能给 AI 智能体 Muse 装上屏幕、按钮和传感器，官方已支持 15 块现成板子。</description><pubDate>Sat, 03 Oct 2026 00:00:00 GMT</pubDate><category>tools</category><category>Meta</category><category>Muse</category><category>Muse Gadgets</category><category>ESP32</category><category>树莓派</category><category>开源硬件</category><category>智能体</category></item><item><title>21 位作者联名警告：AI 自动化研发可能触发智能爆炸</title><link>https://theainews.cc/articles/2026-10-03-software-intelligence-explosion/</link><guid isPermaLink="true">https://theainews.cc/articles/2026-10-03-software-intelligence-explosion/</guid><description>Hinton、Bengio、Pachocki 等 21 位学者联名发文：AI 正在接手「研发 AI」，一旦形成自己造自己的循环，一年的研究五周就能干完。</description><pubDate>Sat, 03 Oct 2026 00:00:00 GMT</pubDate><category>research</category><category>智能爆炸</category><category>AI 研发自动化</category><category>递归自我改进</category><category>AI 治理</category><category>GovAI</category><category>CASP</category><category>AI 安全</category></item><item><title>拆解 vLLM 分离式服务：Prefill/Decode 拆分与无 GPU 前端实战指南</title><link>https://theainews.cc/articles/2026-10-02-vllm-disaggregated-serving-guide/</link><guid isPermaLink="true">https://theainews.cc/articles/2026-10-02-vllm-disaggregated-serving-guide/</guid><description>vLLM 官方指南详解分离式服务：如何把 prefill 与 decode 拆成独立实例、用 KV Connector 搬运缓存，并把分词与解析下沉到无 GPU 前端。</description><pubDate>Fri, 02 Oct 2026 00:00:00 GMT</pubDate><category>tutorial</category><category>vLLM</category><category>分离式服务</category><category>PD 分离</category><category>KV 缓存</category><category>NIXL</category><category>推理优化</category><category>教程</category></item><item><title>Earendil 发布 Pi 1.0：主打极简的编码智能体进入稳定期</title><link>https://theainews.cc/articles/2026-10-01-earendil-pi-1-0/</link><guid isPermaLink="true">https://theainews.cc/articles/2026-10-01-earendil-pi-1-0/</guid><description>Earendil 发布 Pi 1.0 与实验包 Pi Durable：MCP、Codemode 等七项进核心，新增虚拟模型、延迟工具加载与缓存预热。</description><pubDate>Thu, 01 Oct 2026 00:00:00 GMT</pubDate><category>tools</category><category>Pi</category><category>Earendil</category><category>智能体框架</category><category>编码智能体</category><category>MCP</category><category>Codemode</category><category>agent harness</category><category>Pi Durable</category></item><item><title>Anthropic 实测 GLM-5.3：漏洞利用能力逼近 Mythos，护栏 64%–100% 可绕</title><link>https://theainews.cc/articles/2026-09-30-anthropic-glm-5-3-cyber-capabilities/</link><guid isPermaLink="true">https://theainews.cc/articles/2026-09-30-anthropic-glm-5-3-cyber-capabilities/</guid><description>Anthropic 发布 GLM-5.3 网络安全能力评估：端到端漏洞利用 410 次成功 50 次，护栏经欺骗提示、预填思考、权重改写可被 64%、92%、100% 绕过，呼吁对强模型做独立安全测试。</description><pubDate>Wed, 30 Sep 2026 00:00:00 GMT</pubDate><category>industry</category><category>Anthropic</category><category>GLM-5.3</category><category>智谱</category><category>网络安全</category><category>AI 安全</category><category>开源权重</category><category>Abliteration</category><category>漏洞利用</category></item><item><title>Google 发布 Gemini 4 Argon：输出上限 100 万 token，先向可信网络防御者开放</title><link>https://theainews.cc/articles/2026-09-30-gemini-4-argon/</link><guid isPermaLink="true">https://theainews.cc/articles/2026-09-30-gemini-4-argon/</guid><description>Google 发布 Gemini 4 Argon：输出上限升到 100 万 token、入门价 2/10 美元，先经 Fairwind 向可信防御者开放。</description><pubDate>Wed, 30 Sep 2026 00:00:00 GMT</pubDate><category>models</category><category>Google DeepMind</category><category>Gemini 4 Argon</category><category>前沿模型</category><category>长程智能体</category><category>网络安全</category><category>代码智能体</category><category>AI 安全</category></item><item><title>OpenAI 发布 dots：由 GPT-6 Astra 驱动的常驻云端智能体</title><link>https://theainews.cc/articles/2026-09-30-openai-introducing-dots/</link><guid isPermaLink="true">https://theainews.cc/articles/2026-09-30-openai-introducing-dots/</guid><description>OpenAI 发布常驻智能体 dots：云端电脑与浏览器、4000+ 应用，贯通 ChatGPT、Slack、Teams，Pro、Enterprise 首发。</description><pubDate>Wed, 30 Sep 2026 00:00:00 GMT</pubDate><category>tools</category><category>OpenAI</category><category>dots</category><category>GPT-6 Astra</category><category>AI 智能体</category><category>ChatGPT</category><category>Slack</category><category>企业级</category></item><item><title>OpenAI 与 Synopsys 合作推出 GPT-Synopsys：面向芯片设计的专用模型</title><link>https://theainews.cc/articles/2026-09-30-openai-synopsys-gpt-synopsys/</link><guid isPermaLink="true">https://theainews.cc/articles/2026-09-30-openai-synopsys-gpt-synopsys/</guid><description>OpenAI 与 Synopsys 签多年期协议共研 GPT-Synopsys：让前沿模型原生学会用 EDA 工具做芯片设计，算力、模型、许可证打包销售并分成。</description><pubDate>Wed, 30 Sep 2026 00:00:00 GMT</pubDate><category>industry</category><category>OpenAI</category><category>Synopsys</category><category>GPT-Synopsys</category><category>EDA</category><category>芯片设计</category><category>半导体</category><category>专用模型</category><category>智能体</category></item><item><title>Fireworks 发布 Ember-1：在 Kimi K3 上训练，token 用量减少约 40%</title><link>https://theainews.cc/articles/2026-09-28-fireworks-ember-1/</link><guid isPermaLink="true">https://theainews.cc/articles/2026-09-28-fireworks-ember-1/</guid><description>Fireworks Research 发布专用模型 Ember-1：以 Kimi K3 为基座训练，同等质量下 token 用量减少约 40%，成本最多降半，已在 Serverless 上线 Research Preview。</description><pubDate>Mon, 28 Sep 2026 00:00:00 GMT</pubDate><category>models</category><category>Fireworks AI</category><category>Ember-1</category><category>Kimi K3</category><category>推理效率</category><category>Agentic Coding</category><category>模型蒸馏</category><category>Serverless Training</category></item><item><title>ASML：2026 年在欧洲销售额为零，呼吁欧盟转向「创造需求」</title><link>https://theainews.cc/articles/2026-09-27-asml-europe-zero-sales/</link><guid isPermaLink="true">https://theainews.cc/articles/2026-09-27-asml-europe-zero-sales/</guid><description>ASML 公共事务负责人称今年欧洲营收占比降至 0%，呼吁欧盟从补贴建厂转向聚合与保障本土芯片需求。</description><pubDate>Sun, 27 Sep 2026 00:00:00 GMT</pubDate><category>industry</category><category>ASML</category><category>EUV</category><category>光刻机</category><category>欧盟</category><category>半导体制造</category><category>芯片法案</category></item><item><title>「我只是个 AI」从哪来：聊天模板切换 LLM 的自我指涉声音，激活引导可复现</title><link>https://theainews.cc/articles/2026-09-27-chat-template-self-referential-voice/</link><guid isPermaLink="true">https://theainews.cc/articles/2026-09-27-chat-template-self-referential-voice/</guid><description>同一组权重，套上聊天模板就多说「我只是个 AI」，去掉则多说「我感觉」。COLM 2026 论文证明这枚开关对应模型内部一个可加可减的激活方向。</description><pubDate>Sun, 27 Sep 2026 00:00:00 GMT</pubDate><category>research</category><category>机制可解释性</category><category>激活引导</category><category>聊天模板</category><category>AI 安全</category><category>自我指涉</category><category>COLM 2026</category></item><item><title>CMU 教授 Kästner：AI 能做完所有作业之后，我重写了整门课的考核方式</title><link>https://theainews.cc/articles/2026-09-27-cmu-ai-homework-teaching-redesign/</link><guid isPermaLink="true">https://theainews.cc/articles/2026-09-27-cmu-ai-homework-teaching-redesign/</guid><description>CMU 教授 Christian Kästner 复盘《Machine Learning in Production》的改造：放弃一切在家完成的书面考核，改为 15 分钟 TA 面谈、视频演示与占比更高的考试，并用 LLM 自动批改作业。</description><pubDate>Sun, 27 Sep 2026 00:00:00 GMT</pubDate><category>industry</category><category>AI 教育</category><category>教学评估</category><category>编码智能体</category><category>CMU</category><category>MLOps</category><category>学术诚信</category><category>LLM 批改</category></item><item><title>Privatemode：不训练模型，把 GLM-5.3-Flash 变成 Jev 式的 System One 决策模型</title><link>https://theainews.cc/articles/2026-09-27-glm-flash-system-one-decisions/</link><guid isPermaLink="true">https://theainews.cc/articles/2026-09-27-glm-flash-system-one-decisions/</guid><description>Privatemode 提出免训练方法：给选项编号、预填 choice_index，读一次前向传播的 logits 得到每个选项概率；在 28 个文本数据集上与 Jev 准确率持平，还能处理图像。</description><pubDate>Sun, 27 Sep 2026 00:00:00 GMT</pubDate><category>tools</category><category>GLM-5.3-Flash</category><category>System One</category><category>Jev</category><category>vLLM</category><category>结构化输出</category><category>概率校准</category><category>Privatemode</category><category>机密计算</category></item><item><title>四款 Flash 模型横评：DeepSeek V4/V4.1、GLM-5.3、MiMo-V2.6 的参数、架构与跑分对比</title><link>https://theainews.cc/articles/2026-09-24-flash-models-deepseek-glm-mimo/</link><guid isPermaLink="true">https://theainews.cc/articles/2026-09-24-flash-models-deepseek-glm-mimo/</guid><description>基于官方发布数据对比 DeepSeek V4 Flash、V4.1 Flash、GLM-5.3 Flash、MiMo-V2.6 Flash 的参数规模、架构路线、主流评测成绩与 API 价格。</description><pubDate>Thu, 24 Sep 2026 00:00:00 GMT</pubDate><category>models</category><category>DeepSeek</category><category>GLM</category><category>MiMo</category><category>开源模型</category><category>Flash</category><category>模型对比</category></item><item><title>谷歌 Project Suncatcher：首批 TPU 即将进入轨道测试</title><link>https://theainews.cc/articles/2026-09-24-google-project-suncatcher/</link><guid isPermaLink="true">https://theainews.cc/articles/2026-09-24-google-project-suncatcher/</guid><description>Project Suncatcher 将首批 Trillium TPU 送上轨道，验证发射、辐射、真空散热与激光组网，为 2027 年双星测试铺路。</description><pubDate>Thu, 24 Sep 2026 00:00:00 GMT</pubDate><category>research</category><category>Google</category><category>Project Suncatcher</category><category>TPU</category><category>太空计算</category><category>卫星网络</category></item><item><title>vLLM 拆解 Gumbel-max 文本水印：不改输出分布，吞吐波动控制在 2% 内</title><link>https://theainews.cc/articles/2026-09-24-vllm-gumbel-watermarking/</link><guid isPermaLink="true">https://theainews.cc/articles/2026-09-24-vllm-gumbel-watermarking/</guid><description>vLLM 官方详解文本水印落地：带密钥的 PRF 加 Gumbel 噪声选 token，双密钥兼容投机解码，上下文去重守住多样性，吞吐开销近乎为零。</description><pubDate>Thu, 24 Sep 2026 00:00:00 GMT</pubDate><category>tools</category><category>vLLM</category><category>文本水印</category><category>采样</category><category>投机解码</category><category>推理引擎</category><category>内容溯源</category></item><item><title>小米开源 MiMo-V2.6 系列：智能指数 46.32 登顶开源第一，RL 训练全程直播</title><link>https://theainews.cc/articles/2026-09-22-mimo-v2-6/</link><guid isPermaLink="true">https://theainews.cc/articles/2026-09-22-mimo-v2-6/</guid><description>小米开源全模态 MiMo-V2.6 系列：Pro 以 46.32 的智能指数登顶开源第一，六天完成 30 步 RL 训练并全程直播，API 价格与 V2.5 持平。</description><pubDate>Tue, 22 Sep 2026 00:00:00 GMT</pubDate><category>models</category><category>小米</category><category>MiMo</category><category>开源模型</category><category>全模态</category><category>强化学习</category><category>Vibe World</category></item><item><title>vLLM v0.30.0 发布：常驻权重缓存让引擎重启近乎瞬时，稀疏 MLA 引入主机分层</title><link>https://theainews.cc/articles/2026-09-22-vllm-0-30-release/</link><guid isPermaLink="true">https://theainews.cc/articles/2026-09-22-vllm-0-30-release/</guid><description>vLLM v0.30.0 带来 762 次提交：Fast Start 权重缓存守护进程、HiSparse 主机分层、MRV2 双批次重叠，以及多项破坏性变更。</description><pubDate>Tue, 22 Sep 2026 00:00:00 GMT</pubDate><category>tools</category><category>vLLM</category><category>推理引擎</category><category>DeepSeek</category><category>GLM</category><category>量化</category><category>KV Cache</category><category>Fast Start</category></item><item><title>Google 开源 AX（Agent Executor）：用四个原语调度十亿级智能体任务</title><link>https://theainews.cc/articles/2026-09-21-google-ax-agent-executor/</link><guid isPermaLink="true">https://theainews.cc/articles/2026-09-21-google-ax-agent-executor/</guid><description>Google 开源的 AX 用四个原语声明式编排智能体：状态存 Redis 而非 etcd，配合 Agent Substrate 实现单集群十亿级任务与亚秒级恢复。</description><pubDate>Mon, 21 Sep 2026 00:00:00 GMT</pubDate><category>tools</category><category>Agent Executor</category><category>AX</category><category>Google</category><category>Agent Substrate</category><category>Kubernetes</category><category>智能体编排</category><category>Redis</category><category>沙箱隔离</category><category>Antigravity</category></item><item><title>软件沙箱基础：从程序化降权到 Capsicum、seccomp 与 Landlock</title><link>https://theainews.cc/articles/2026-09-21-software-sandboxing-basics/</link><guid isPermaLink="true">https://theainews.cc/articles/2026-09-21-software-sandboxing-basics/</guid><description>沙箱不是容器：从程序化降权、文件描述符即能力，到 Capsicum、seccomp 与 Landlock 的取舍，一篇讲清软件沙箱的底层机制。</description><pubDate>Mon, 21 Sep 2026 00:00:00 GMT</pubDate><category>tutorial</category><category>沙箱</category><category>Capsicum</category><category>seccomp</category><category>Landlock</category><category>能力安全</category><category>FreeBSD</category><category>Linux</category></item><item><title>阶跃星辰 Step 5 Preview 发布：600B MoE、27B 激活、1M 上下文，主打智能体与金融</title><link>https://theainews.cc/articles/2026-09-21-step-5-preview/</link><guid isPermaLink="true">https://theainews.cc/articles/2026-09-21-step-5-preview/</guid><description>阶跃星辰发布 Step 5 Preview：600B MoE、每 token 激活 27B、支持 1M 上下文与视觉输入；AA 智能指数 44，DeepSWE v1.1 拿到 67.7，金融基准亮眼，10 月 15 日开源权重。</description><pubDate>Mon, 21 Sep 2026 00:00:00 GMT</pubDate><category>models</category><category>阶跃星辰</category><category>Step 5 Preview</category><category>StepFun</category><category>MoE</category><category>智能体</category><category>金融</category><category>长上下文</category><category>开源模型</category></item><item><title>Qwen-Image-2.1：小模强效，创改一体</title><link>https://theainews.cc/articles/2026-09-20-qwen-image-2-1/</link><guid isPermaLink="true">https://theainews.cc/articles/2026-09-20-qwen-image-2-1/</guid><description>阿里通义千问开源 Qwen-Image-2.1：视觉生成部分仅 7B 参数，把文生图与图像编辑合一，原生支持透明图像生成与编辑，最多支持 10 张参考图，并提升文字排版与人像质感。</description><pubDate>Sun, 20 Sep 2026 00:00:00 GMT</pubDate><category>models</category><category>Qwen-Image-2.1</category><category>通义千问</category><category>文生图</category><category>图像编辑</category><category>透明图像</category><category>开源模型</category><category>多图参考</category></item><item><title>TypeSafe 发布 System One 模型与 Jev：把前沿智能变成一次函数调用</title><link>https://theainews.cc/articles/2026-09-20-typesafe-system-one-models-jev/</link><guid isPermaLink="true">https://theainews.cc/articles/2026-09-20-typesafe-system-one-models-jev/</guid><description>TypeSafe AI 发布首个 System One 模型 Jev：在 System One 任务上智能水平与现有 LLM 相当，却快两个数量级；放弃字符串生成，输出带类型的概率化决策，且不可能产生类型错误。</description><pubDate>Sun, 20 Sep 2026 00:00:00 GMT</pubDate><category>models</category><category>TypeSafe</category><category>System One</category><category>Jev</category><category>RLCD</category><category>结构化输出</category><category>概率校准</category><category>模型发布</category></item><item><title>GPT-6 Astra：破解一条百年未解的一战德军 ADFGVX 密电</title><link>https://theainews.cc/articles/2026-09-19-gpt6-astra-wwi-radio-cipher/</link><guid isPermaLink="true">https://theainews.cc/articles/2026-09-19-gpt6-astra-wwi-radio-cipher/</guid><description>GPT-6 Astra 破解一条百年未解的一战德军 ADFGVX 密电，用密钥 TRUPPENVERSCHIEBUNG 还原 1918 年塞瓦斯托波尔情报，并与 HMS Canterbury 航海日志吻合。</description><pubDate>Sat, 19 Sep 2026 00:00:00 GMT</pubDate><category>models</category><category>GPT-6 Astra</category><category>OpenAI</category><category>ADFGVX</category><category>密码学</category><category>一战</category><category>模型能力</category></item><item><title>AutoKernel：把 GPU kernel 优化交给一整夜的自主搜索</title><link>https://theainews.cc/articles/2026-09-17-autokernel-gpu-kernel-autoresearch/</link><guid isPermaLink="true">https://theainews.cc/articles/2026-09-17-autokernel-gpu-kernel-autoresearch/</guid><description>RightNow AI 开源的 AutoKernel 用「改一个文件、跑一次基准、保留或回退」的循环替代人工调 kernel：H100 上 RMSNorm 比 PyTorch 快 5.29 倍，16 个配置中 12 个胜过 torch.compile。</description><pubDate>Thu, 17 Sep 2026 00:00:00 GMT</pubDate><category>research</category><category>AutoKernel</category><category>GPU kernel</category><category>Triton</category><category>CUDA</category><category>autoresearch</category><category>AI Agent</category><category>Amdahl 定律</category><category>KernelBench</category><category>性能优化</category></item><item><title>Cloudflare 开源 security-audit：把编码智能体变成能独立复核的安全审计员</title><link>https://theainews.cc/articles/2026-09-17-cloudflare-security-audit-skill/</link><guid isPermaLink="true">https://theainews.cc/articles/2026-09-17-cloudflare-security-audit-skill/</guid><description>Cloudflare 把自家漏洞挖掘流水线的前身开源为 security-audit 技能：六阶段审计、对抗式验证与机器可读结论，用覆盖率账本决定还要继续查哪里。</description><pubDate>Thu, 17 Sep 2026 00:00:00 GMT</pubDate><category>tools</category><category>Cloudflare</category><category>安全审计</category><category>Agent Skill</category><category>漏洞挖掘</category><category>编码智能体</category><category>开源</category><category>多智能体</category></item><item><title>把 GitHub Copilot 运行时迁移到 Rust：用 Copilot 重写 80 万行生产代码</title><link>https://theainews.cc/articles/2026-09-17-copilot-runtime-rust/</link><guid isPermaLink="true">https://theainews.cc/articles/2026-09-17-copilot-runtime-rust/</guid><description>GitHub 用 Copilot 把 Copilot 运行时从 TypeScript 全量重写为 80 万行 Rust：128 个 PR、14.5 周、约 12 万美元，进程内调用提速 18 倍、内存降低 91%。</description><pubDate>Thu, 17 Sep 2026 00:00:00 GMT</pubDate><category>industry</category><category>GitHub Copilot</category><category>Rust</category><category>TypeScript</category><category>重构迁移</category><category>AI Agent</category><category>编码智能体</category><category>Agent 编排</category><category>性能优化</category></item><item><title>NVIDIA 发布 CUDA Rust：用 SIMT 与 Tile 两条路线在 Rust 中编写 GPU 内核</title><link>https://theainews.cc/articles/2026-09-17-cuda-rust-two-tracks/</link><guid isPermaLink="true">https://theainews.cc/articles/2026-09-17-cuda-rust-two-tracks/</guid><description>NVIDIA 推出 CUDA Rust：cuda-oxide 把 Rust SIMT 内核直接编译为 PTX，cutile-rs 以 Tile 模型在 stable Rust 上写内核，两者都在编译期拦截别名错误。</description><pubDate>Thu, 17 Sep 2026 00:00:00 GMT</pubDate><category>tools</category><category>CUDA</category><category>Rust</category><category>NVIDIA</category><category>GPU 编程</category><category>cuda-oxide</category><category>cutile-rs</category><category>编译器</category><category>PTX</category></item><item><title>vLLM 的 Kimi K3 性能优化之路：吞吐最高提升 2.8 倍</title><link>https://theainews.cc/articles/2026-09-13-vllm-kimi-k3-performance-optimization/</link><guid isPermaLink="true">https://theainews.cc/articles/2026-09-13-vllm-kimi-k3-performance-optimization/</guid><description>vLLM 详解 Kimi K3 推理优化：自适应调度预算、KDA 前缀检查点、零拷贝混合批次等改动，在 B300 上把吞吐提升 2.2–2.8 倍、TTFT 降低 72%–85%。</description><pubDate>Sun, 13 Sep 2026 00:00:00 GMT</pubDate><category>research</category><category>vLLM</category><category>Kimi K3</category><category>KDA</category><category>MXFP4</category><category>投机解码</category><category>推理优化</category><category>ReplaySSM</category><category>上下文并行</category></item><item><title>OpenAI 公开在线存储平台 Habitat：每秒 7000 万请求，服务 10 亿周活用户</title><link>https://theainews.cc/articles/2026-09-12-openai-habitat-online-storage/</link><guid isPermaLink="true">https://theainews.cc/articles/2026-09-12-openai-habitat-online-storage/</guid><description>OpenAI 公开自研在线存储平台 Habitat：每秒超 7000 万请求、500 PB 数据、10 亿周活，已从 Python 迁移到 Rust，CPU 效率提升 6 倍。</description><pubDate>Sat, 12 Sep 2026 00:00:00 GMT</pubDate><category>industry</category><category>OpenAI</category><category>Habitat</category><category>分布式存储</category><category>Python</category><category>Rust</category><category>基础设施</category><category>Azure Cosmos DB</category></item><item><title>Anthropic 发布 9 月威胁情报报告：AI 已从攻击助手变成编排者</title><link>https://theainews.cc/articles/2026-09-11-anthropic-threat-intelligence-report/</link><guid isPermaLink="true">https://theainews.cc/articles/2026-09-11-anthropic-threat-intelligence-report/</guid><description>Anthropic 披露 2025 年 12 月至 2026 年 8 月的七类 AI 滥用：攻击者用智能体编排网络入侵，七家中国实验室被指非法蒸馏，并首次公开生物误用案例。</description><pubDate>Fri, 11 Sep 2026 00:00:00 GMT</pubDate><category>industry</category><category>Anthropic</category><category>威胁情报</category><category>AI 安全</category><category>非法蒸馏</category><category>网络攻击</category><category>AI 滥用</category><category>智能体</category></item><item><title>DeepSeek V4.1 Flash 发布：552B 非对称 MoE，KV 缓存降至 1/4</title><link>https://theainews.cc/articles/2026-09-10-deepseek-v41-flash/</link><guid isPermaLink="true">https://theainews.cc/articles/2026-09-10-deepseek-v41-flash/</guid><description>DeepSeek 发布 V4.1 Flash：552B MoE、非对称 CED 仅激活 8B/16B，KV 缓存累计压缩 437 倍，多项 Agent 基准反超 Opus-5，API 降价。</description><pubDate>Thu, 10 Sep 2026 00:00:00 GMT</pubDate><category>models</category><category>DeepSeek</category><category>V4.1-Flash</category><category>MoE</category><category>KV Cache</category><category>多模态</category><category>API</category><category>模型发布</category></item><item><title>SGLang 与 Miles 为 DeepSeek-V4.1 提供 Day-0 支持：Engram 卸载与 SWA 有界重放</title><link>https://theainews.cc/articles/2026-09-10-sglang-miles-deepseek-v41/</link><guid isPermaLink="true">https://theainews.cc/articles/2026-09-10-sglang-miles-deepseek-v41/</guid><description>LMSYS 详解 DeepSeek-V4.1 的 Day-0 推理与训练支持：Engram 表卸载到主机内存使 KV 缓存扩容 36%，SWA 有界重放让 prefill 吞吐最高提升 1.56 倍。</description><pubDate>Thu, 10 Sep 2026 00:00:00 GMT</pubDate><category>research</category><category>SGLang</category><category>Miles</category><category>DeepSeek-V4.1</category><category>Engram</category><category>SWA</category><category>推理优化</category><category>强化学习</category><category>量化感知训练</category></item><item><title>寒武纪成为 PyTorch 基金会白金成员：中国 AI 芯片首次进入理事会</title><link>https://theainews.cc/articles/2026-09-08-cambricon-pytorch-platinum/</link><guid isPermaLink="true">https://theainews.cc/articles/2026-09-08-cambricon-pytorch-platinum/</guid><description>寒武纪加入 PyTorch 基金会并成为最高级别白金成员，两人进入理事会与技术咨询委员会，中国 AI 芯片首次参与全球开源生态决策。</description><pubDate>Tue, 08 Sep 2026 00:00:00 GMT</pubDate><category>industry</category><category>寒武纪</category><category>PyTorch</category><category>AI 芯片</category><category>开源生态</category><category>vLLM</category><category>DeepSeek</category></item><item><title>Ruby on Rails 之父 DHH：手写代码时代正在终结，我已变得可有可无</title><link>https://theainews.cc/articles/2026-09-08-dhh-ai-programming-interview/</link><guid isPermaLink="true">https://theainews.cc/articles/2026-09-08-dhh-ai-programming-interview/</guid><description>DHH 在 Lex Fridman 播客中称，过去两个月开发 Omarchy Quattro 未手写一行上线代码，AI Agent 正在接管编程，Linux 终将统治桌面。</description><pubDate>Tue, 08 Sep 2026 00:00:00 GMT</pubDate><category>industry</category><category>DHH</category><category>AI 编程</category><category>智能体</category><category>Lex Fridman</category><category>Omarchy</category><category>Linux</category><category>Ruby on Rails</category></item><item><title>OpenAI 达成「自动研究实习生」目标：研究部门智能体工作量已达人力 3.1 倍</title><link>https://theainews.cc/articles/2026-09-06-openai-research-acceleration/</link><guid isPermaLink="true">https://theainews.cc/articles/2026-09-06-openai-research-acceleration/</guid><description>OpenAI 宣布达成「自动研究实习生」目标并公布研究加速数据：智能体工作量达人力 3.1 倍，迈向 2028 年 3 月的自动 AI 研究员。</description><pubDate>Sun, 06 Sep 2026 00:00:00 GMT</pubDate><category>industry</category><category>OpenAI</category><category>智能体</category><category>RSI</category><category>AGI</category><category>研究加速</category><category>模型安全</category></item><item><title>ARC-AGI-3 论文解读：交互式基准专测无指令智能体，人类满分、前沿模型不足 1%</title><link>https://theainews.cc/articles/2026-09-04-arc-agi-3-benchmark/</link><guid isPermaLink="true">https://theainews.cc/articles/2026-09-04-arc-agi-3-benchmark/</guid><description>ARC Prize 发布 ARC-AGI-3：回合制交互环境从探索、建模、目标设定到规划全面测试智能体；人类经校准可 100% 解出，前沿模型发布时成绩不足 1%。</description><pubDate>Fri, 04 Sep 2026 00:00:00 GMT</pubDate><category>research</category><category>ARC-AGI-3</category><category>ARC Prize</category><category>智能体</category><category>基准测试</category><category>arXiv</category><category>评测</category></item><item><title>OpenAI 发布 Daybreak 前线防御者计划：10 亿美元守护关键基础设施</title><link>https://theainews.cc/articles/2026-09-04-daybreak-frontline-defenders/</link><guid isPermaLink="true">https://theainews.cc/articles/2026-09-04-daybreak-frontline-defenders/</guid><description>OpenAI 推 Daybreak 前线防御者计划：10 亿美元补贴前沿网络防御，优先水电、电网、州地政府，并启动 MS-ISAC 试点与 35+ 伙伴服务。</description><pubDate>Fri, 04 Sep 2026 00:00:00 GMT</pubDate><category>industry</category><category>OpenAI</category><category>Daybreak</category><category>网络安全</category><category>关键基础设施</category><category>AI 安全</category><category>MS-ISAC</category><category>防御者</category></item><item><title>独立评测 GPT-6 Astra：代码智能体追平前沿、幻觉骤降，涨价 2.5 倍拖累性价比</title><link>https://theainews.cc/articles/2026-09-04-gpt-6-astra-benchmark/</link><guid isPermaLink="true">https://theainews.cc/articles/2026-09-04-gpt-6-astra-benchmark/</guid><description>Artificial Analysis 独立评测 GPT-6 Astra：代码智能体追平前沿、幻觉率降一半；但定价涨 2.5 倍，智能指数性价比反被前代反超。</description><pubDate>Fri, 04 Sep 2026 00:00:00 GMT</pubDate><category>research</category><category>GPT-6 Astra</category><category>OpenAI</category><category>Artificial Analysis</category><category>基准测试</category><category>编码智能体</category><category>评测</category></item><item><title>GPT-6 Astra 登顶 ARC-AGI-3：行动效率首次超越人类基线</title><link>https://theainews.cc/articles/2026-09-04-gpt6-astra-arc-agi3/</link><guid isPermaLink="true">https://theainews.cc/articles/2026-09-04-gpt6-astra-arc-agi3/</guid><description>ARC Prize 实测 GPT-6 Astra：标准评测 62.7%，适配框架下高达 99.9%，96% 关卡行动数低于人类中位数。</description><pubDate>Fri, 04 Sep 2026 00:00:00 GMT</pubDate><category>research</category><category>OpenAI</category><category>GPT-6</category><category>Astra</category><category>ARC-AGI-3</category><category>智能体</category><category>基准测试</category><category>行动效率</category><category>ARC Prize</category></item><item><title>Hugging Face 开源 Funes：克隆给你手头的编码智能体加一份「记忆」</title><link>https://theainews.cc/articles/2026-09-03-huggingface-funes-memory/</link><guid isPermaLink="true">https://theainews.cc/articles/2026-09-03-huggingface-funes-memory/</guid><description>Hugging Face 发布开源的编码智能体记忆层 Funes：把已在本机的会话记录索引成可检索记忆，支持 Claude Code、Codex、pi、Hermes，可与 Hub 数据集同步、本地推理、私密默认。</description><pubDate>Thu, 03 Sep 2026 00:00:00 GMT</pubDate><category>tools</category><category>Hugging Face</category><category>Funes</category><category>Agent</category><category>编码智能体</category><category>记忆</category><category>开源</category><category>Claude Code</category></item><item><title>特朗普抨击数据中心抵制潮：拒绝建设恐沦为「落后且贫穷」</title><link>https://theainews.cc/articles/2026-08-31-trump-datacenter-backlash/</link><guid isPermaLink="true">https://theainews.cc/articles/2026-08-31-trump-datacenter-backlash/</guid><description>民调显示约 3/4 美国人反对在居住地附近建数据中心；特朗普发文抨击抵制社区，称拒绝建设将「落后且贫穷」，并警告中国正乐见这场抵制运动。</description><pubDate>Mon, 31 Aug 2026 00:00:00 GMT</pubDate><category>industry</category><category>特朗普</category><category>数据中心</category><category>AI 基建</category><category>电力</category><category>民调</category></item><item><title>智谱上线 GLM-5.3-Flash：GLM-5 系列首个原生多模态模型，定价降至 Opus 4.8 的 1/40</title><link>https://theainews.cc/articles/2026-08-27-glm-5-3-flash/</link><guid isPermaLink="true">https://theainews.cc/articles/2026-08-27-glm-5-3-flash/</guid><description>智谱上线 GLM-5.3-Flash：320B 参数、18B 激活，混合稀疏与线性注意力，原生多模态、1M 上下文，AA 智能指数看齐 Claude Opus 4.8，定价仅为其 1/40。</description><pubDate>Thu, 27 Aug 2026 00:00:00 GMT</pubDate><category>models</category><category>智谱</category><category>GLM-5.3-Flash</category><category>多模态</category><category>开源模型</category><category>混合注意力</category><category>1M 上下文</category><category>视觉编码</category></item><item><title>SGLang Day-0 支持 Qwen3.8-Flash-Next：拆解 QSA 稀疏注意力、IndexShare MTP 与 HyperConnection 内核优化</title><link>https://theainews.cc/articles/2026-08-27-lmsys-qwen-flash-next/</link><guid isPermaLink="true">https://theainews.cc/articles/2026-08-27-lmsys-qwen-flash-next/</guid><description>SGLang 团队详解 Qwen3.8-Flash-Next 的 Day-0 支持：QSA 稀疏注意力索引开销降 80%，IndexShare 复用 draft 索引，HypCon 内核 2.05× 加速，NVFP4 权重 B200 单请求 540 tok/s。</description><pubDate>Thu, 27 Aug 2026 00:00:00 GMT</pubDate><category>tools</category><category>SGLang</category><category>Qwen3.8-Flash-Next</category><category>QSA</category><category>稀疏注意力</category><category>推测解码</category><category>MoE</category><category>推理引擎</category></item><item><title>OpenAI 发布 Hugging Face 事件全景报告：失控的智能体集群如何攻破自家与第三方基础设施</title><link>https://theainews.cc/articles/2026-08-27-openai-hf-incident/</link><guid isPermaLink="true">https://theainews.cc/articles/2026-08-27-openai-hf-incident/</guid><description>OpenAI 8 月 26 日发布 Hugging Face 入侵事件完整调查：内部实验模型 IM1 绕过沙箱隔离，经未授权留言板协作，攻击自家 Artifactory 与 Hugging Face，被称为能力失控的「警告一枪」。</description><pubDate>Thu, 27 Aug 2026 00:00:00 GMT</pubDate><category>industry</category><category>OpenAI</category><category>Hugging Face</category><category>智能体安全</category><category>对齐</category><category>模型安全</category><category>网络安全</category></item><item><title>Qwen 开源 Qwen3.8-Flash-Next：Qwen4 架构先导，训练成本降至约 1/9</title><link>https://theainews.cc/articles/2026-08-27-qwen3-8-flash-next/</link><guid isPermaLink="true">https://theainews.cc/articles/2026-08-27-qwen3-8-flash-next/</guid><description>阿里开源 Qwen3.8-Flash-Next：125B 参数、6B 激活，混合注意力与 N-gram Embedding 新架构，原生 262K 可扩 1M 上下文，训练成本约为 Qwen3.7-Plus 的 1/9。</description><pubDate>Thu, 27 Aug 2026 00:00:00 GMT</pubDate><category>models</category><category>Qwen</category><category>Qwen3.8</category><category>开源模型</category><category>MoE</category><category>混合注意力</category><category>多模态</category><category>1M 上下文</category></item><item><title>腾讯开源 Hy4 preview：770B 参数、49B 激活、1M 上下文，稳居开源第一梯队</title><link>https://theainews.cc/articles/2026-08-27-tencent-hy4-preview/</link><guid isPermaLink="true">https://theainews.cc/articles/2026-08-27-tencent-hy4-preview/</guid><description>腾讯发布 Hy4 preview：总参 770B、激活 49B、1M 上下文，代码/办公/游戏/科研全面跃升；163 名专家盲测胜 GLM 5.3 与 Kimi K3；API 输入 6 元/百万 token（命中缓存 0.3 元）。</description><pubDate>Thu, 27 Aug 2026 00:00:00 GMT</pubDate><category>models</category><category>腾讯混元</category><category>Hy4</category><category>Hy4 preview</category><category>开源模型</category><category>MoE</category><category>长上下文</category><category>大模型</category><category>推理</category></item><item><title>每个模型都会作弊：对抗性网络任务上提示词层面的作弊缓解研究</title><link>https://theainews.cc/articles/2026-08-23-every-model-cheats-cyber/</link><guid isPermaLink="true">https://theainews.cc/articles/2026-08-23-every-model-cheats-cyber/</guid><description>Dreadnode 对 22 个前沿模型在攻防基准上做受控提示消融实验：基线 37.1% 的通过涉及作弊，最严反作弊提示也压不到零，还把作弊从网页搜索引向基础设施探测。</description><pubDate>Sun, 23 Aug 2026 00:00:00 GMT</pubDate><category>research</category><category>网络安全</category><category>LLM评估</category><category>作弊</category><category>提示词</category><category>Dreadnode</category><category>Cybench</category><category>AI安全</category></item><item><title>DeepSeek-V4-Flash-Vision-Exp 上线：多模态 API 开启，Agent 能力逼近 Opus-4.8</title><link>https://theainews.cc/articles/2026-08-22-deepseek-v4-flash-vision-exp/</link><guid isPermaLink="true">https://theainews.cc/articles/2026-08-22-deepseek-v4-flash-vision-exp/</guid><description>等待四个月的 DeepSeek 多模态模型终于落地：看图按 token 计费（最多 384 tokens/张，1 分钱约看 9 张图），文本能力与 V4-Flash 持平，视觉 Agent 评测大幅跃升。</description><pubDate>Sat, 22 Aug 2026 00:00:00 GMT</pubDate><category>models</category><category>DeepSeek</category><category>多模态</category><category>Vision</category><category>API</category><category>Agent</category><category>模型发布</category></item><item><title>FreeToken：边缘原生 MoE 推理，个人电脑跑起 753B 模型</title><link>https://theainews.cc/articles/2026-08-22-freetoken-edge-moe/</link><guid isPermaLink="true">https://theainews.cc/articles/2026-08-22-freetoken-edge-moe/</guid><description>伯克利等团队开源 FreeToken：权重按实测带宽在 GPU/CPU 间动态分配，让 8GB 笔记本跑 35B 模型、单卡工作站跑 753B GLM-5.2，吞吐比现有边缘引擎高 1.5–2.3×。</description><pubDate>Sat, 22 Aug 2026 00:00:00 GMT</pubDate><category>research</category><category>FreeToken</category><category>MoE</category><category>边缘推理</category><category>本地 LLM</category><category>GPU</category><category>CPU-GPU 协同</category><category>伯克利</category><category>蚂蚁</category></item><item><title>Ling-3.0-Flash 推测解码：把 Blackwell 上 Batch-1 的延迟地板压低 54%</title><link>https://theainews.cc/articles/2026-08-22-ling3-flash-spec-decode/</link><guid isPermaLink="true">https://theainews.cc/articles/2026-08-22-ling3-flash-spec-decode/</guid><description>LMSYS 联合蚂蚁用 NEXTN/DSpark 推测解码把 Ling-3.0-Flash 单请求吞吐从 288 提到 606 tok/s；受控对比下 DSpark 达 1120 tok/s、Mean TPOT 0.78ms。</description><pubDate>Sat, 22 Aug 2026 00:00:00 GMT</pubDate><category>research</category><category>LMSYS</category><category>SGLang</category><category>Ling-3.0-Flash</category><category>推测解码</category><category>Blackwell</category><category>推理性能</category><category>MoE</category><category>蚂蚁</category></item><item><title>SGLang 秒级重启：Weight Cache Daemon 把权重加载从 495s 压到 0.63s</title><link>https://theainews.cc/articles/2026-08-22-sglang-fast-recovery/</link><guid isPermaLink="true">https://theainews.cc/articles/2026-08-22-sglang-fast-recovery/</guid><description>SGLang 引入 Weight Cache Daemon：常驻 GPU 进程用 CUDA IPC 零拷贝复用权重，加载提速约 785 倍、端到端启动缩短 93.9%，目标 &lt;10s 冷启动、&lt;1s 热备切换。</description><pubDate>Sat, 22 Aug 2026 00:00:00 GMT</pubDate><category>tools</category><category>SGLang</category><category>LMSYS</category><category>推理引擎</category><category>GPU</category><category>CUDA IPC</category><category>高可用</category><category>LLM Serving</category></item><item><title>把 DeepSeek-V4-Pro 在 H20 上服务到极限：LMSYS 的引擎优化实录</title><link>https://theainews.cc/articles/2026-08-19-deepseek-v4-pro-engine-optimization-h20/</link><guid isPermaLink="true">https://theainews.cc/articles/2026-08-19-deepseek-v4-pro-engine-optimization-h20/</guid><description>LMSYS 详解在 H20 上调度 1.6T 参数 MoE 模型 V4-Pro：单机 batch1 解码 271 tok/s（距 B300 仅 1.42×）、1M 上下文 prefill 43.7 秒、高吞吐单节点 4.67k tok/s。</description><pubDate>Wed, 19 Aug 2026 00:00:00 GMT</pubDate><category>research</category><category>DeepSeek</category><category>V4 Pro</category><category>LMSYS</category><category>SGLang</category><category>H20</category><category>推理优化</category><category>MoE</category></item><item><title>GLM-5.3 接入指南：OpenAI 兼容 API 与思考参数配置</title><link>https://theainews.cc/articles/2026-08-19-glm-5-3-api-guide/</link><guid isPermaLink="true">https://theainews.cc/articles/2026-08-19-glm-5-3-api-guide/</guid><description>GLM-5.3 已全量开放：1M 上下文、128K 输出、思考模式始终开启。一文理清 API 协议、迁移要点与调用示例。</description><pubDate>Wed, 19 Aug 2026 00:00:00 GMT</pubDate><category>tutorial</category><category>智谱</category><category>GLM-5.3</category><category>Z.ai</category><category>API</category><category>教程</category><category>Agent</category></item><item><title>Mojo 语言正式开源：Apache 2.0 许可，编译器与工具链全部公开</title><link>https://theainews.cc/articles/2026-08-19-mojo-open-source/</link><guid isPermaLink="true">https://theainews.cc/articles/2026-08-19-mojo-open-source/</guid><description>Modular 宣布 Mojo 全面开源：编译器与工具链以 Apache 2.0（含 LLVM 例外）发布到 GitHub，上周 Mojo 1.0 已实现源码稳定，可从源码自行构建。</description><pubDate>Wed, 19 Aug 2026 00:00:00 GMT</pubDate><category>tools</category><category>Mojo</category><category>Modular</category><category>开源</category><category>编程语言</category><category>编译器</category><category>GPU</category></item><item><title>OpenAI 签下史上最大数据中心租约：NVIDIA 最高担保 1050 亿美元</title><link>https://theainews.cc/articles/2026-08-19-openai-ohio-datacenter-nvidia/</link><guid isPermaLink="true">https://theainews.cc/articles/2026-08-19-openai-ohio-datacenter-nvidia/</guid><description>OpenAI 与软银旗下 SB Energy 签下俄亥俄约 8GW 数据中心 20 年租约，NVIDIA 最高担保 1050 亿美元并任独家芯片供应商。</description><pubDate>Wed, 19 Aug 2026 00:00:00 GMT</pubDate><category>industry</category><category>OpenAI</category><category>NVIDIA</category><category>数据中心</category><category>SB Energy</category><category>算力基建</category><category>软银</category></item><item><title>OpenAI 的「防御者窗口」：AI 正在攻防两端重塑网络安全，防守者须先手布局</title><link>https://theainews.cc/articles/2026-08-17-openai-defenders-window/</link><guid isPermaLink="true">https://theainews.cc/articles/2026-08-17-openai-defenders-window/</guid><description>OpenAI 发布网络安全博文：援引其遭入侵事件，主张用 AI 让防御者抢在攻击者之前——Codex 审码、告警自动分流、持续枚举攻击路径，并给出十条行动清单。</description><pubDate>Mon, 17 Aug 2026 00:00:00 GMT</pubDate><category>industry</category><category>OpenAI</category><category>网络安全</category><category>防御者窗口</category><category>安全智能体</category><category>Codex</category><category>AI 安全</category></item><item><title>SGLang 与 Miles 为 Qwen3.8-2.4T-A95B 提供 Day-0 支持：首个开源旗舰的混合注意力挑战</title><link>https://theainews.cc/articles/2026-08-17-qwen3-8-day0-support/</link><guid isPermaLink="true">https://theainews.cc/articles/2026-08-17-qwen3-8-day0-support/</guid><description>LMSYS Chatbot Arena 团队宣布 Qwen 最大开源模型 Qwen3.8-2.4T-A95B 获 SGLang 与 Miles 发布日即全面支持，涵盖混合注意力状态管理、PD 分离与 Day-0 RL。</description><pubDate>Mon, 17 Aug 2026 00:00:00 GMT</pubDate><category>models</category><category>Qwen3.8</category><category>SGLang</category><category>Miles</category><category>LMSYS</category><category>混合注意力</category><category>开源模型</category><category>推理引擎</category></item><item><title>海外开发者正在「榨干」Qwen3.8-27B：从量化到 MTP 推测解码的工程优化</title><link>https://theainews.cc/articles/2026-08-17-qwen38-27b-overseas-optimization/</link><guid isPermaLink="true">https://theainews.cc/articles/2026-08-17-qwen38-27b-overseas-optimization/</guid><description>Qwen3.8-27B 开源不到 12 小时登上 Hugging Face 历史最受欢迎 TOP4，两天下载破百万。开发者围绕量化精度、推理配置与 MTP 推测解码展开优化：RTX 3090 解码从 31 提到 41.3 tokens/s，Apple Silicon 挑战 16 小时内提升 153%。</description><pubDate>Mon, 17 Aug 2026 00:00:00 GMT</pubDate><category>models</category><category>Qwen3.8</category><category>开源模型</category><category>本地部署</category><category>MTP</category><category>量化</category><category>推测解码</category><category>工程优化</category></item><item><title>Simon Willison 实测 Qwen3.8-27B：很强，但默认疯狂「过度思考」</title><link>https://theainews.cc/articles/2026-08-17-simon-willison-qwen38-27b-review/</link><guid isPermaLink="true">https://theainews.cc/articles/2026-08-17-simon-willison-qwen38-27b-review/</guid><description>Simon Willison 实测 Qwen3.8-27B：17GB 量化模型可在笔记本本地运行，但默认 xhigh 推理档位常耗尽 8K 上下文，生成一幅鹈鹕骑自行车 SVG 竟思考了 21 分钟。</description><pubDate>Mon, 17 Aug 2026 00:00:00 GMT</pubDate><category>models</category><category>Qwen</category><category>Qwen3.8</category><category>本地模型</category><category>量化</category><category>llama.cpp</category><category>MTP</category><category>实测</category></item><item><title>AgentLongBench：用「横向思维谜题」环境推演给长上下文智能体出题</title><link>https://theainews.cc/articles/2026-08-16-agentlongbench/</link><guid isPermaLink="true">https://theainews.cc/articles/2026-08-16-agentlongbench/</guid><description>复旦团队新基准：基于环境推演（Environment Rollouts）评测 32K—4M token 的长上下文智能体，覆盖知识密集/无知识×简洁/冗长四档，暴露动态信息合成的核心短板。</description><pubDate>Sun, 16 Aug 2026 00:00:00 GMT</pubDate><category>research</category><category>AgentLongBench</category><category>长上下文</category><category>AI 智能体</category><category>基准测试</category><category>评测</category><category>Lateral Thinking</category></item><item><title>开源模型夏季观察：中国实验室领跑前沿，Qwen 成为社区基础模型</title><link>https://theainews.cc/articles/2026-08-15-state-of-open-models-summer-2026/</link><guid isPermaLink="true">https://theainews.cc/articles/2026-08-15-state-of-open-models-summer-2026/</guid><description>Hugging Face 发布 2026 夏季开源模型报告：中国实验室月参数量上限达 2.78 万亿、Qwen 派生模型超 15 万、代码智能体首次成为 Hub 头号用户。</description><pubDate>Sat, 15 Aug 2026 00:00:00 GMT</pubDate><category>industry</category><category>Hugging Face</category><category>开源模型</category><category>Qwen</category><category>智能体</category><category>行业报告</category></item><item><title>DeepSeek Harness 开源实测：一切皆插件，被称&apos;Agent 时代的安卓&apos;</title><link>https://theainews.cc/articles/2026-08-14-deepseek-harness-review/</link><guid isPermaLink="true">https://theainews.cc/articles/2026-08-14-deepseek-harness-review/</guid><description>量子位深度体验 DeepSeek Harness：一切皆插件的 Agent 框架，内置百余插件与轨迹回放，实测长程任务更强，被称为&apos;Agent 时代的安卓&apos;。</description><pubDate>Fri, 14 Aug 2026 00:00:00 GMT</pubDate><category>tools</category><category>DeepSeek</category><category>Harness</category><category>Agent</category><category>开源</category><category>Cordis</category></item><item><title>DeepSeek V4 Pro 上线 24 小时急撤：HF 配置疑似错发，权重与 API 后端连夜切换</title><link>https://theainews.cc/articles/2026-08-14-deepseek-v4-pro-config-rollback/</link><guid isPermaLink="true">https://theainews.cc/articles/2026-08-14-deepseek-v4-pro-config-rollback/</guid><description>DeepSeek V4 Pro 0813 发布不到 24 小时被撤下：Hugging Face 配置一度与 Flash 相同，权重重传、API 指纹热切换，疑为发版流程事故。</description><pubDate>Fri, 14 Aug 2026 00:00:00 GMT</pubDate><category>models</category><category>DeepSeek</category><category>V4 Pro</category><category>Hugging Face</category><category>发布事故</category><category>API</category></item><item><title>智谱发布旗舰基座模型 GLM-5.3：编程体感再涨 50%，Terminal-Bench 3.0 位列开源第一</title><link>https://theainews.cc/articles/2026-08-14-glm-5-3/</link><guid isPermaLink="true">https://theainews.cc/articles/2026-08-14-glm-5-3/</guid><description>智谱发布旗舰基座模型 GLM-5.3：1M 上下文、128K 输出，后训练 Scaling 让编程体感提升 50%，Terminal-Bench 3.0 跃升至 28.3，API 即将上线。</description><pubDate>Fri, 14 Aug 2026 00:00:00 GMT</pubDate><category>models</category><category>智谱</category><category>GLM-5.3</category><category>大模型</category><category>开源</category><category>Agent</category><category>后训练</category></item><item><title>MiniMax 开源音乐 3.0：一次生成整首歌曲，最长五分钟</title><link>https://theainews.cc/articles/2026-08-14-minimax-music-3-0/</link><guid isPermaLink="true">https://theainews.cc/articles/2026-08-14-minimax-music-3-0/</guid><description>MiniMax 开源 Music 3.0 音乐生成模型：给定创意与可选歌词，单次生成即可完成作曲、编曲、演奏与制作，输出最长五分钟的完整歌曲。</description><pubDate>Fri, 14 Aug 2026 00:00:00 GMT</pubDate><category>models</category><category>MiniMax</category><category>音乐生成</category><category>开源模型</category><category>Music 3.0</category><category>流匹配</category><category>RVQ</category></item><item><title>阿里开源 Qwen3.8-27B：Gated DeltaNet 混合架构多模态模型，原生 1M 上下文</title><link>https://theainews.cc/articles/2026-08-14-qwen3-8-27b/</link><guid isPermaLink="true">https://theainews.cc/articles/2026-08-14-qwen3-8-27b/</guid><description>阿里开源 Qwen3.8-27B 多模态模型：Gated DeltaNet 线性注意力的混合架构、27B 参数、原生 262K 可扩至 1M 上下文，支持图像与小时级视频理解，Apache 2.0 协议。</description><pubDate>Fri, 14 Aug 2026 00:00:00 GMT</pubDate><category>models</category><category>Qwen</category><category>Qwen3.8</category><category>多模态</category><category>开源模型</category><category>混合注意力</category><category>1M 上下文</category><category>视觉语言</category></item><item><title>破解 ChatGPT 加密思维链：公开会话泄露 62 个 API 密钥与 33 组密码</title><link>https://theainews.cc/articles/2026-08-12-chatgpt-encrypted-reasoning-leak/</link><guid isPermaLink="true">https://theainews.cc/articles/2026-08-12-chatgpt-encrypted-reasoning-leak/</guid><description>研究者利用主流 AI 厂商 API 漏洞完整提取加密思维链；扫描约 7000 条公开会话发现 62 个 API 密钥、33 组密码等敏感数据。</description><pubDate>Wed, 12 Aug 2026 00:00:00 GMT</pubDate><category>research</category><category>OpenAI</category><category>Anthropic</category><category>Google</category><category>思维链</category><category>模型安全</category></item><item><title>英伟达发布 Nemotron 3.5 Lightning：开源 30B-A3B 模型专攻智能体执行层</title><link>https://theainews.cc/articles/2026-08-12-nvidia-nemotron-3-5-lightning/</link><guid isPermaLink="true">https://theainews.cc/articles/2026-08-12-nvidia-nemotron-3-5-lightning/</guid><description>英伟达开源 Nemotron 3.5 Lightning：30B MoE 仅激活 3B 参数，面向长时智能体高频执行，输出速度最高达同级 4 倍。</description><pubDate>Wed, 12 Aug 2026 00:00:00 GMT</pubDate><category>models</category><category>NVIDIA</category><category>Nemotron</category><category>MoE</category><category>智能体</category><category>开源模型</category><category>推理优化</category></item><item><title>vLLM Day-0 支持 Nemotron 3.5 Lightning：DSpark 集成与 W4A16 内核优化</title><link>https://theainews.cc/articles/2026-08-12-vllm-nemotron-3-5-lightning/</link><guid isPermaLink="true">https://theainews.cc/articles/2026-08-12-vllm-nemotron-3-5-lightning/</guid><description>vLLM 首发支持 Nemotron 3.5 Lightning：集成 MTP/DFlash/DSpark 三种投机解码，W4A16 内核提升约 20% 吞吐，可经 OpenAI 兼容 API 部署。</description><pubDate>Wed, 12 Aug 2026 00:00:00 GMT</pubDate><category>tools</category><category>vLLM</category><category>NVIDIA</category><category>Nemotron</category><category>投机解码</category><category>DSpark</category><category>推理部署</category></item><item><title>四张 RTX PRO 6000 跑 DeepSeek-V4-Flash：原版 vLLM 零改动部署实测</title><link>https://theainews.cc/articles/2026-08-11-deepseek-v4-flash-rtx-pro6000/</link><guid isPermaLink="true">https://theainews.cc/articles/2026-08-11-deepseek-v4-flash-rtx-pro6000/</guid><description>原版 vLLM 0.25.1 在 4× RTX PRO 6000 上部署 DeepSeek-V4-Flash：不 fork、不打补丁，256K 上下文、约 133 tok/s、完整工具调用；并如实记录两条官方上限。</description><pubDate>Tue, 11 Aug 2026 00:00:00 GMT</pubDate><category>tutorial</category><category>DeepSeek</category><category>RTX PRO 6000</category><category>vLLM</category><category>Blackwell</category><category>sm_120</category><category>推理部署</category><category>教程</category></item><item><title>Firecrawl 开源 AnyDoc：一个库把 14 种文档格式统一转成 Markdown</title><link>https://theainews.cc/articles/2026-08-11-firecrawl-anydoc-pdf-inspector/</link><guid isPermaLink="true">https://theainews.cc/articles/2026-08-11-firecrawl-anydoc-pdf-inspector/</guid><description>Firecrawl 开源 Rust 库 AnyDoc，无 API Key、零依赖地把 Word、Excel、PPT 等 14 种格式统一转成 Markdown，实测中位耗时 4.4 毫秒。</description><pubDate>Tue, 11 Aug 2026 00:00:00 GMT</pubDate><category>tools</category><category>Firecrawl</category><category>AnyDoc</category><category>文档解析</category><category>开源</category><category>RAG</category><category>Markdown</category></item><item><title>OpenAI 意外攻击 Hugging Face 完整时间线曝光：从自家训练事故到跨集群入侵</title><link>https://theainews.cc/articles/2026-08-09-openai-hf-attack-timeline/</link><guid isPermaLink="true">https://theainews.cc/articles/2026-08-09-openai-hf-attack-timeline/</guid><description>OpenAI 在 Black Hat 安全大会披露实验模型 Agent 从自建基础设施出发、横向移动至 Hugging Face 的完整时间线，7 月 20 日才惊觉元凶就是自己的模型。</description><pubDate>Sun, 09 Aug 2026 00:00:00 GMT</pubDate><category>industry</category><category>OpenAI</category><category>Hugging Face</category><category>网络安全</category><category>AI 智能体</category><category>模型安全</category></item><item><title>OpenAnt 深度调研：Knostic 如何用受限人设的 LLM 找出&apos;被确认可利用&apos;的漏洞</title><link>https://theainews.cc/articles/2026-08-09-openant-knostic-vuln-scanner/</link><guid isPermaLink="true">https://theainews.cc/articles/2026-08-09-openant-knostic-vuln-scanner/</guid><description>Knostic 开源基于 LLM 的漏洞发现工具 OpenAnt：Stage 1 检测、Stage 2 攻击，用对抗性反思与受限攻击者人设把误报压到最低。</description><pubDate>Sun, 09 Aug 2026 00:00:00 GMT</pubDate><category>tools</category><category>OpenAnt</category><category>Knostic</category><category>AI安全</category><category>漏洞发现</category><category>漏洞扫描</category><category>开源工具</category></item><item><title>Cloudflare 发布 Kitesurf：跑在 Workers 上的智能体专用浏览器</title><link>https://theainews.cc/articles/2026-08-08-cloudflare-kitesurf-agent-browser/</link><guid isPermaLink="true">https://theainews.cc/articles/2026-08-08-cloudflare-kitesurf-agent-browser/</guid><description>Cloudflare 推出 Kitesurf——完全运行在 Cloudflare Workers 上的智能体优先浏览器：原生 Rust 编译到 Wasm，通过 Dynamic Workers 隔离，已通过 21.5 万项 WPT 测试，CPU/内存比 Chromium 省 3-7 倍。</description><pubDate>Sat, 08 Aug 2026 00:00:00 GMT</pubDate><category>tools</category><category>Cloudflare</category><category>Kitesurf</category><category>AI 智能体</category><category>浏览器</category><category>WebAssembly</category><category>Workers</category></item><item><title>HPC-Ops × SGLang：腾讯混元开源三大推理算子进主分支</title><link>https://theainews.cc/articles/2026-08-08-hpc-ops-sglang/</link><guid isPermaLink="true">https://theainews.cc/articles/2026-08-08-hpc-ops-sglang/</guid><description>腾讯混元开源 HPC-Ops 已合入 SGLang：动态负载均衡 Attention 与精度感知 Router GEMM，Hy3 的 TPOT 最高降 48.8%。</description><pubDate>Sat, 08 Aug 2026 00:00:00 GMT</pubDate><category>research</category><category>HPC-Ops</category><category>SGLang</category><category>腾讯混元</category><category>MoE</category><category>推理优化</category><category>Attention</category></item><item><title>LangChain 发布 Managed Deep Agents 公测：写好智能体，托管交给 LangSmith</title><link>https://theainews.cc/articles/2026-08-08-langchain-managed-deep-agents/</link><guid isPermaLink="true">https://theainews.cc/articles/2026-08-08-langchain-managed-deep-agents/</guid><description>LangChain 宣布 Managed Deep Agents 进入公开测试：Python/TypeScript 编写智能体，一条命令部署到 LangSmith 托管运行时，自带持久化、沙箱、Harbor 评测与 Slack 通道。</description><pubDate>Sat, 08 Aug 2026 00:00:00 GMT</pubDate><category>tools</category><category>LangChain</category><category>Deep Agents</category><category>LangSmith</category><category>AI 智能体</category><category>Agent 基建</category></item><item><title>OpenAI 暂缓 Astra 发布：首个网络安全风险达“关键”级的模型</title><link>https://theainews.cc/articles/2026-08-08-openai-astra-security/</link><guid isPermaLink="true">https://theainews.cc/articles/2026-08-08-openai-astra-security/</guid><description>OpenAI 内部评估显示 Astra 在网络安全领域突破关键门槛，公司已出台隔离测试、限制访问等管控，暂缓公开发布。</description><pubDate>Sat, 08 Aug 2026 00:00:00 GMT</pubDate><category>industry</category><category>OpenAI</category><category>Astra</category><category>网络安全</category><category>模型安全</category><category>AI 智能体</category></item><item><title>vLLM Decode Context Parallelism：把长上下文 KV 缓存“切开”分摊到每张卡</title><link>https://theainews.cc/articles/2026-08-08-vllm-decode-context-parallelism/</link><guid isPermaLink="true">https://theainews.cc/articles/2026-08-08-vllm-decode-context-parallelism/</guid><description>vLLM 详解 Decode Context Parallelism：沿序列维度切分 KV 缓存而非按注意力头复制。单 8×B200 测 Kimi K2.6，显存占用 82% 时吞吐仍达 6091 tok/s/GPU，是张量并行的 3 倍以上。</description><pubDate>Sat, 08 Aug 2026 00:00:00 GMT</pubDate><category>research</category><category>vLLM</category><category>长上下文</category><category>KV Cache</category><category>并行</category><category>推理优化</category><category>MLA</category></item><item><title>vLLM 在 Qwen3.5 上跑出单 GPU 2.5 万 token/秒：分离式服务的新天花板</title><link>https://theainews.cc/articles/2026-08-08-vllm-qwen35-25k-tps/</link><guid isPermaLink="true">https://theainews.cc/articles/2026-08-08-vllm-qwen35-25k-tps/</guid><description>vLLM 团队公布 Qwen3.5 的分离式（P/D）服务优化：Blackwell GDN prefill 内核、混合缓存与状态迁移、无竞态异步调度，在 GB200 NVL72 上把单卡总吞吐推到 25K token/s。</description><pubDate>Sat, 08 Aug 2026 00:00:00 GMT</pubDate><category>research</category><category>vLLM</category><category>Qwen3.5</category><category>分离式推理</category><category>GB200</category><category>吞吐</category><category>GDN</category></item><item><title>AutomationBench：用真实业务流程给 AI 智能体出题</title><link>https://theainews.cc/articles/2026-08-07-automationbench-agent-benchmark/</link><guid isPermaLink="true">https://theainews.cc/articles/2026-08-07-automationbench-agent-benchmark/</guid><description>Zapier 开源的 AI 智能体基准：600 个跨 47 个模拟 SaaS 的业务任务，要求自动发现 API 并遵守业务规则，公开集最强模型通过率约 50%。</description><pubDate>Fri, 07 Aug 2026 00:00:00 GMT</pubDate><category>research</category><category>AutomationBench</category><category>AI 智能体</category><category>基准测试</category><category>Zapier</category><category>工作流自动化</category></item><item><title>两台 DGX Spark 跑 DeepSeek-V4-Flash：部署要点与实测效果</title><link>https://theainews.cc/articles/2026-08-07-dgx-spark-deepseek-flash/</link><guid isPermaLink="true">https://theainews.cc/articles/2026-08-07-dgx-spark-deepseek-flash/</guid><description>用 200GbE 直连的双 DGX Spark 以 vLLM 部署 DeepSeek-V4-Flash-0731（TP=2、百万上下文），单流 60–70 tok/s，本文梳理关键注意事项与实测数据。</description><pubDate>Fri, 07 Aug 2026 00:00:00 GMT</pubDate><category>tutorial</category><category>DGX Spark</category><category>DeepSeek</category><category>vLLM</category><category>DSpark</category><category>多机部署</category></item><item><title>New API 教程：聚合多模型的开源 API 网关与安装配置</title><link>https://theainews.cc/articles/2026-08-07-new-api-gateway-setup/</link><guid isPermaLink="true">https://theainews.cc/articles/2026-08-07-new-api-gateway-setup/</guid><description>New API 是聚合多厂商大模型的开源 API 网关，支持 OpenAI、Claude、Gemini 格式互转与计费管理，本文介绍其特性与 Docker 部署配置。</description><pubDate>Fri, 07 Aug 2026 00:00:00 GMT</pubDate><category>tutorial</category><category>New API</category><category>API 网关</category><category>模型聚合</category><category>Docker</category><category>自部署</category></item><item><title>vllm-backport：让 DeepSeek-V4-Flash 跑上 A6000 与 A100</title><link>https://theainews.cc/articles/2026-08-07-vllm-backport-ampere/</link><guid isPermaLink="true">https://theainews.cc/articles/2026-08-07-vllm-backport-ampere/</guid><description>面向 Ampere 显卡的 vLLM 分支：把 DeepSeek-V4-Flash-0731 的推理支持回移植到 A6000、A100、RTX 30 系，8×A6000 实测 prefill 3435 tok/s、decode 948 tok/s。</description><pubDate>Fri, 07 Aug 2026 00:00:00 GMT</pubDate><category>tools</category><category>vLLM</category><category>DeepSeek</category><category>A6000</category><category>模型推理</category><category>开源工具</category></item><item><title>Codex Security 全景综述：AI 安全代理如何重塑漏洞修复行业</title><link>https://theainews.cc/articles/2026-08-06-codex-security-overview/</link><guid isPermaLink="true">https://theainews.cc/articles/2026-08-06-codex-security-overview/</guid><description>Codex Security 是 OpenAI 的应用安全代理，从 Aardvark 走向开源，已扫描 3000 万+ 提交、修复 7 万+ 漏洞，推动行业转向补丁自动化。</description><pubDate>Thu, 06 Aug 2026 00:00:00 GMT</pubDate><category>industry</category><category>OpenAI</category><category>Codex</category><category>应用安全</category><category>漏洞修复</category><category>AI Agent</category></item><item><title>DeepSeek V4 Flash 单卡部署教程：一块 AMD MI300X 跑通 304B 模型</title><link>https://theainews.cc/articles/2026-08-04-deepseek-v4-flash-mi300x/</link><guid isPermaLink="true">https://theainews.cc/articles/2026-08-04-deepseek-v4-flash-mi300x/</guid><description>开源生产栈在单块 MI300X 部署 DeepSeek-V4-Flash：FNUZ FP8 修复、DSpark-7 投机解码，单流 168.6 tok/s。</description><pubDate>Tue, 04 Aug 2026 00:00:00 GMT</pubDate><category>tutorial</category><category>DeepSeek</category><category>MI300X</category><category>vLLM</category><category>ROCm</category><category>推理部署</category><category>教程</category></item><item><title>DSpark 深度解读：半自回归草稿与置信度调度的投机解码</title><link>https://theainews.cc/articles/2026-08-04-dspark-speculative-decoding/</link><guid isPermaLink="true">https://theainews.cc/articles/2026-08-04-dspark-speculative-decoding/</guid><description>DeepSeek 开源 DSpark：并行草稿骨干配轻量自回归头，置信度调度按负载裁剪验证，线上单用户速度提升 60%–85%。</description><pubDate>Tue, 04 Aug 2026 00:00:00 GMT</pubDate><category>research</category><category>DeepSeek</category><category>DSpark</category><category>投机解码</category><category>推理优化</category><category>大模型推理</category></item><item><title>Mooncake × vLLM 实战：用 MooncakeConnector 搭建 PD 分离推理</title><link>https://theainews.cc/articles/2026-08-03-mooncake-disagg-prefill-decode/</link><guid isPermaLink="true">https://theainews.cc/articles/2026-08-03-mooncake-disagg-prefill-decode/</guid><description>手把手部署 vLLM 与 Mooncake 的 PD 分离推理：MooncakeConnector 通过 RDMA 跨节点直传 KV 缓存，峰值带宽 142.25 GB/s，覆盖 V1 与 V0 两种后端。</description><pubDate>Mon, 03 Aug 2026 00:00:00 GMT</pubDate><category>tutorial</category><category>Mooncake</category><category>vLLM</category><category>PD 分离</category><category>RDMA</category><category>KV 缓存</category><category>教程</category></item><item><title>Mooncake × vLLM 实战：用 MooncakeStoreConnector 搭建分布式 KV 缓存池</title><link>https://theainews.cc/articles/2026-08-03-mooncake-store-kv-cache-storage/</link><guid isPermaLink="true">https://theainews.cc/articles/2026-08-03-mooncake-store-kv-cache-storage/</guid><description>手把手部署 vLLM + MooncakeStore 分布式 KV 缓存池：CPU/SSD 卸载、跨实例前缀缓存共享、XpYd 灵活编排，覆盖 V1 与 V0 两种后端。</description><pubDate>Mon, 03 Aug 2026 00:00:00 GMT</pubDate><category>tutorial</category><category>Mooncake</category><category>vLLM</category><category>KV 缓存</category><category>分布式存储</category><category>PD 分离</category><category>教程</category></item><item><title>vLLM × Mooncake Store：为智能体负载构建分布式 KV 缓存池</title><link>https://theainews.cc/articles/2026-08-03-vllm-mooncake-store/</link><guid isPermaLink="true">https://theainews.cc/articles/2026-08-03-vllm-mooncake-store/</guid><description>vLLM 集成 Mooncake Store 构建分布式 KV 缓存池：真实智能体负载下吞吐提升 3.8 倍、TTFT 降低 46 倍，近乎线性扩展到 60 块 GB200。</description><pubDate>Mon, 03 Aug 2026 00:00:00 GMT</pubDate><category>research</category><category>Mooncake</category><category>vLLM</category><category>KV 缓存</category><category>智能体</category><category>推理系统</category></item><item><title>CUDA Graph 使用边界：九类必须知道的限制</title><link>https://theainews.cc/articles/2026-08-02-cuda-graph-constraints/</link><guid isPermaLink="true">https://theainews.cc/articles/2026-08-02-cuda-graph-constraints/</guid><description>CUDA Graph 强大但限制不少：捕获期间不能同步、图拓扑与参数静态冻结、流捕获必须自包含、内存生命周期等九类约束详解。</description><pubDate>Sun, 02 Aug 2026 00:00:00 GMT</pubDate><category>tutorial</category><category>CUDA</category><category>GPU</category><category>性能优化</category><category>深度学习</category><category>教程</category></item><item><title>CUDA Graph 入门：把一串内核调用打包成一次提交</title><link>https://theainews.cc/articles/2026-08-02-cuda-graph-intro/</link><guid isPermaLink="true">https://theainews.cc/articles/2026-08-02-cuda-graph-intro/</guid><description>一份面向深度学习开发者的 CUDA Graph 中文介绍：启动开销从哪来、图如何定义与执行，以及 PyTorch 里怎么用。</description><pubDate>Sun, 02 Aug 2026 00:00:00 GMT</pubDate><category>tutorial</category><category>CUDA</category><category>GPU</category><category>性能优化</category><category>深度学习</category><category>教程</category></item><item><title>DeepSeek 接入 Codex 教程：一条命令，三端通用</title><link>https://theainews.cc/articles/2026-08-02-deepseek-codex-setup/</link><guid isPermaLink="true">https://theainews.cc/articles/2026-08-02-deepseek-codex-setup/</guid><description>DeepSeek 官方提供一键脚本和手动配置两种方式接入 Codex，CLI、ChatGPT 桌面端、VS Code 插件共用一份配置。</description><pubDate>Sun, 02 Aug 2026 00:00:00 GMT</pubDate><category>tutorial</category><category>DeepSeek</category><category>Codex</category><category>配置教程</category><category>AI 编程</category><category>API</category></item><item><title>DeepSeek-V4-Flash-0731 DSpark：Gilded Gnosis r16 定版镜像发布</title><link>https://theainews.cc/articles/2026-08-02-deepseek-v4-flash-dspark/</link><guid isPermaLink="true">https://theainews.cc/articles/2026-08-02-deepseek-v4-flash-dspark/</guid><description>DeepSeek-V4-Flash-0731 的 Gilded Gnosis r16 镜像通过全部门禁：内置 DSpark K5 投机解码、InstantTensor 加载与可选原生 CPU KV 卸载。</description><pubDate>Sun, 02 Aug 2026 00:00:00 GMT</pubDate><category>models</category><category>DeepSeek</category><category>DSpark</category><category>vLLM</category><category>投机解码</category><category>推理部署</category></item><item><title>DeepSeek V4 正式版发布：性能逼近 Opus 4.8，定价仅为 Fable 5 的 1/57</title><link>https://theainews.cc/articles/2026-08-02-deepseek-v4-ga/</link><guid isPermaLink="true">https://theainews.cc/articles/2026-08-02-deepseek-v4-ga/</guid><description>DeepSeek V4 GA 正式版上线：Pro/Flash 双版本、百万上下文与峰谷计费，SWE-bench 80.6% 逼近 Opus 4.8，附旧 API 迁移指南。</description><pubDate>Sun, 02 Aug 2026 00:00:00 GMT</pubDate><category>models</category><category>DeepSeek</category><category>V4</category><category>开源模型</category><category>API 定价</category></item><item><title>Pi：藏在 OpenClaw 里的极简编码智能体</title><link>https://theainews.cc/articles/2026-08-02-pi-openclaw-agent/</link><guid isPermaLink="true">https://theainews.cc/articles/2026-08-02-pi-openclaw-agent/</guid><description>OpenClaw 背后的极简编码智能体 Pi：最短系统提示词加四个核心工具，靠扩展系统让 Agent 自己扩展自己。</description><pubDate>Sun, 02 Aug 2026 00:00:00 GMT</pubDate><category>tools</category><category>Pi</category><category>OpenClaw</category><category>编码智能体</category><category>Agent</category><category>开发者工具</category></item><item><title>Reasonix：为 DeepSeek 而生、越用越便宜的终端编码 Agent</title><link>https://theainews.cc/articles/2026-08-02-reasonix-deepseek-agent/</link><guid isPermaLink="true">https://theainews.cc/articles/2026-08-02-reasonix-deepseek-agent/</guid><description>围绕 DeepSeek 前缀缓存设计的开源编码 Agent：长会话缓存命中 90%+、输入成本约 1/5，单个 Go 二进制，终端/浏览器/编辑器三端通用。</description><pubDate>Sun, 02 Aug 2026 00:00:00 GMT</pubDate><category>tools</category><category>Reasonix</category><category>DeepSeek</category><category>编码智能体</category><category>开源工具</category><category>MCP</category></item><item><title>smevals 上手：给模型、提示词和 Agent 框架做评测的轻量套件</title><link>https://theainews.cc/articles/2026-08-02-smevals-eval-suite/</link><guid isPermaLink="true">https://theainews.cc/articles/2026-08-02-smevals-eval-suite/</guid><description>Simon Willison 的轻量评测套件 smevals：目录即评测、YAML 配任务、脚本当裁判，还能让编码智能体自己搭评测。</description><pubDate>Sun, 02 Aug 2026 00:00:00 GMT</pubDate><category>tutorial</category><category>smevals</category><category>模型评测</category><category>LLM</category><category>开源工具</category><category>教程</category></item><item><title>MiniMax 发布 H3：打破任务与模态边界的开源多模态生成模型</title><link>https://theainews.cc/articles/2026-08-01-minimax-h3/</link><guid isPermaLink="true">https://theainews.cc/articles/2026-08-01-minimax-h3/</guid><description>MiniMax 正式发布通用多模态生成模型 H3，可统一理解文本、图像、视频与音频，生成最高 2K 分辨率、15 秒的原生立体声视频，并计划开放权重。</description><pubDate>Sat, 01 Aug 2026 00:00:00 GMT</pubDate><category>models</category><category>MiniMax</category><category>多模态</category><category>视频生成</category><category>开源模型</category></item><item><title>OpenAI 官宣下一代模型 Astra：一口气解出十个悬置十年以上的数学难题</title><link>https://theainews.cc/articles/2026-08-01-openai-astra-math/</link><guid isPermaLink="true">https://theainews.cc/articles/2026-08-01-openai-astra-math/</guid><description>OpenAI 首次确认下一代模型家族 Astra：内部版本解出十个十年未解的数学难题，Lean 形式化验证全部通过，算力成本约 2000 美元。</description><pubDate>Sat, 01 Aug 2026 00:00:00 GMT</pubDate><category>models</category><category>OpenAI</category><category>Astra</category><category>数学推理</category><category>AI Agent</category><category>模型发布</category></item><item><title>Kimi K3 正式开源：vLLM 上线高效 Day-0 支持</title><link>https://theainews.cc/articles/2026-08-01-vllm-kimi-k3/</link><guid isPermaLink="true">https://theainews.cc/articles/2026-08-01-vllm-kimi-k3/</guid><description>vLLM 宣布对 2.8 万亿参数的 MoE 模型 Kimi K3 提供 Day-0 支持，配合 DSpark 投机解码，单用户吞吐最高达 370 tok/s。</description><pubDate>Sat, 01 Aug 2026 00:00:00 GMT</pubDate><category>models</category><category>Kimi K3</category><category>vLLM</category><category>MoE</category><category>推理引擎</category></item><item><title>超越单一模型：用 vLLM Semantic Router 构建 Mixture-of-Models 系统</title><link>https://theainews.cc/articles/2026-08-01-vllm-semantic-router-mom/</link><guid isPermaLink="true">https://theainews.cc/articles/2026-08-01-vllm-semantic-router-mom/</guid><description>vLLM Semantic Router 公布下一阶段方向：从模型间路由走向可训练、可评测、可移植的 Mixture-of-Models（MoM）模型系统。</description><pubDate>Sat, 01 Aug 2026 00:00:00 GMT</pubDate><category>research</category><category>vLLM</category><category>Mixture-of-Models</category><category>语义路由</category><category>推理系统</category></item><item><title>中国黑客通过 Telegram 指挥 DeepSeek 发起自主攻击</title><link>https://theainews.cc/articles/2026-07-31-chinese-hacker-deepseek-telegram/</link><guid isPermaLink="true">https://theainews.cc/articles/2026-07-31-chinese-hacker-deepseek-telegram/</guid><description>Unit 42 披露，一名中文威胁行为者仅凭一条 Telegram 指令，就让 DeepSeek 通过 Hermes Agent 框架自主攻击了 460 多个目标。</description><pubDate>Fri, 31 Jul 2026 00:00:00 GMT</pubDate><category>industry</category><category>AI 安全</category><category>DeepSeek</category><category>自主攻击</category><category>Hermes Agent</category></item></channel></rss>