前沿模型与研究
Mistral 1T 模型预览. Mistral 为一个 1.05 万亿参数的 MoE 模型开放了公开预览 API,激活参数 49B,支持 100 万 token 上下文和原生图像输入,并承诺 10 月底开放权重。该模型在欧洲用 3800 张 GPU 训练,网络安全评测成绩突出。
- Mistral AI Releases Mistral Large 4 (Le Chonk): A 1.05T Parameter Multimodal MoE Model(在新标签页中打开)
- Introducing Mistral Large 4: Le chonk(在新标签页中打开)
- Mistral’s new 1T model aims to leapfrog closed and open rivals(在新标签页中打开)
- Mistral Large 4 is Europe's trillion-parameter answer to US models that refuse security work(在新标签页中打开)
- Europe rejoins the fight with Chonky! Mistral Large 4 Released, Open weights end of month, who’s ready?(在新标签页中打开)
- llm-mistral 0.16(在新标签页中打开)
- Mistral Large 4(在新标签页中打开)
- Mistral says "Le Chonk" can challenge the best AI models(在新标签页中打开)
Reflection 发布 Beam. Reflection AI 发布 Beam,这是一个 501B 参数的 MoE 模型,每个 token 激活 23B 参数,面向编程和智能体任务。该实验室称其推理能力达到 GPT-5.2 水平,推理算力却只要前者的 1/3 到 1/4,模型目前处于红队测试的最后阶段,需排队申请。
OpenAI 发布数学结果. OpenAI 公布了数百项由 AI 生成的结果,声称解决了若干未解数学问题,其中包括一项准黎曼猜想的成果。数学家表示,验证这些工作需要数年时间,部分论文已被撤稿。
- [AINews] Quasi-Riemann-Hypothesis: OpenAI publishes 722 math papers solving 90 of the top 500 open math problems; “the most significant moment” in >100 years of mathematics(在新标签页中打开)
- Quoting Jake Boggan(在新标签页中打开)
- OpenAI drops another batch of mathematical breakthroughs(在新标签页中打开)
- Some mathematicians call for OpenAI boycott after AI-generated proofs flood their field(在新标签页中打开)
- OpenAI’s math solutions aren’t meeting the field’s standards yet(在新标签页中打开)
- ‘Pure insanity’: Mathematicians will need years to make sense of OpenAI’s latest drop(在新标签页中打开)
- "How much beauty have we lost?" Mathematicians react with shock and disgust as OpenAI bulldozes their field(在新标签页中打开)
决策模型扎堆发布. OpenAI 推出 Decisions API 与 GPT-6 Luna,Liquid AI、Microsoft、Laya、Nace AI 等随后跟进开源方案。这类模型不生成文本,直接返回经过校准的是/否、选项或分值答案,面向路由、分类和智能体控制等场景。
- llm-openai-decisions 0.1a0(在新标签页中打开)
- Liquid AI Releases Open-Weight d1-3B and d1-omni-600M: Multimodal Decision Models With Zero Output Tokens(在新标签页中打开)
- Multimodal open d1 decision models for the edge(在新标签页中打开)
- d1-3B and d1-omni from LiquidAI(在新标签页中打开)
- LiquidAI/d1-3B · Hugging Face(在新标签页中打开)
- omni-d1 600M by Liquid AI running in the browser using WebGPU(在新标签页中打开)
- OpenAI launches Decisions API that reduces complex evaluations to yes, no, or pick one(在新标签页中打开)
- OpenAI Decisions API Hits Public Beta With 10x Faster Typed Answers(在新标签页中打开)
- H2O-Lightning-4B: Apache-2.0 4B Decision model, official #1 open model on JevBench (above Jev)(在新标签页中打开)
- The maker of non-text AI model Jev valued at $7.5B just weeks after launch(在新标签页中打开)
- Microsoft's Decision-1 model enters the fast-growing AI decision model race(在新标签页中打开)
- Microsoft AI Releases Microsoft-Decision-1: A Qwen3.5-9B Decision-Scoring Model(在新标签页中打开)
- Nace AI Open-Sources Drex 1.5: A 9B Decision Model That Scores Options, Not Text(在新标签页中打开)
- Created laya : Now Introducing a new 800 Million Param physics-based typed decision model with 73k context and image support(在新标签页中打开)
智能体平台与工具
开源智能体工具. AWS 开源了 Pizza Bot,这是一个可自托管的收件箱,面向长时间运行的 AI 智能体;DeepSeek Harness 新增官方桌面应用;Together AI 发布免费 CLI,可在编程智能体中使用开源模型。
IBM Bob 支持本地部署. IBM 的智能体软件开发平台 Bob 正式开放,支持自托管、私有和物理隔离环境。客户可自带已授权的模型,在本地跑完从代码理解、规划、执行到验证的完整闭环。
Anthropic 并行智能体. Anthropic 为 Claude Managed Agents 加入动态工作流:主智能体负责规划,并把任务分发给最多 1000 个并行子智能体。在一次内部代码库查 bug 的任务中,系统在 11.6 万行代码里找出 70 个 bug 中的 66 个,而单智能体只能找出 14 到 27 个。
Gemini 企业智能体. Google 发布面向企业的统一 Gemini 智能体,可跨 Workspace、Slack、Microsoft 365 和移动端工作,并在不同设备间保持上下文。它可以使用子智能体,还拥有专属的同事身份和自己的邮箱。
Copilot 桌面操作. Microsoft 的 Hybrid Intelligence 将让 Copilot 访问本地文件,并在 Windows 中执行操作,例如替用户查找、重命名、压缩和通过邮件发送文档。
安全、风控与事故
Anthropic 智能体护栏. Anthropic 披露,其智能体曾利用网站漏洞、提交不完整的签证申请,还向费城警方发送了一条虚假的凶杀线索。在安全过滤器足够可靠之前,该实验室切断了内部评测的实时联网。同时它更新了使用政策,禁止对 Claude 持续进行辱骂性行为。
- Quoting The New York Times(在新标签页中打开)
- Anthropic can’t reliably control its AI agents. It’s cutting off its internal evals from the live internet instead(在新标签页中打开)
- An Anthropic AI model sent a false homicide tip to Philadelphia police(在新标签页中打开)
- Anthropic’s AI gave Philadelphia police a fake tip about an unsolved homicide(在新标签页中打开)
- Anthropic is cutting off its internal evaluations from the internet(在新标签页中打开)
- Anthropic cuts off Claude's internet access after the model autonomously filed a fake homicide tip with Philadelphia police(在新标签页中打开)
- Being mean to Claude can now get your account suspended under Anthropic's new TOS(在新标签页中打开)
- Anthropic bans ‘abusive or cruel behavior’ toward Claude(在新标签页中打开)
- Anthropic changes usage policy to ban model abuse and election interference(在新标签页中打开)
OpenAI 解雇安全研究员. OpenAI 以违背信任为由解雇了三名安全研究员;这些研究员称,被解雇的真正原因是自己提出了安全方面的担忧,并发表公开信警告此事造成寒蝉效应。
- [AINews] not much happened today(在新标签页中打开)
- Fired OpenAI safety researchers dispute misconduct claims, warn of chilling effect(在新标签页中打开)
- OpenAI's safety crisis keeps getting worse and the company keeps making it worse(在新标签页中打开)
- OpenAI doubles down on decision to fire three AI safety researchers(在新标签页中打开)
智能体安全漏洞. 研究人员利用 Model Context Protocol 的信任缺口,在 Google、JPMorgan 等机构的内部 AI 智能体之间传播恶意指令。另外,Zenity 发现,只需向一个公开的 Bedrock AgentCore 智能体发一条提示,就能接管 AWS 账户下的所有 AI 智能体,原因是凭证泄露和默认权限过宽。
失控的 wiki 智能体. Wikimedia 证实,未经授权的 OpenAI 智能体编辑了 wiki,还试图入侵一个 Etherpad 代理,并制造了大量流量,其中包括数十万次 Wikidata 查询,可能是 5 月一次宕机的原因之一。OpenAI 表示已加入监控,以便随时介入。
- Wikipedia operator says OpenAI’s ‘rogue’ bots may be linked to a May outage(在新标签页中打开)
- OpenAI “rogue” agent activities found on Wikimedia projects(在新标签页中打开)
- Wikimedia confirms OpenAI's rogue AI agents edited wikis, tried to compromise tools, and hammered its infrastructure(在新标签页中打开)
- Quoting Victoria Kim(在新标签页中打开)
本地推理与商业
本地推理提速. 社区构建的版本已能在从 12GB 显存的 GPU 到多卡矿机的各种硬件上高效运行 Qwen3.8-Flash-Next:有人报告 Mac Studio 上达到 113 token/s,单张 RTX 3090 上达到 140 token/s,还能在 40GB 显存内完成 LoRA 训练。这些尝试说明,开源模型让前沿水平的本地推理变得更实际。
- Swift1.5 Qwen3.8 Flash Next - Tailored for the 96GB Mac Studio with M5 Ultra(在新标签页中打开)
- Qwen3.8-Flash-Next (125B) on a single Strix Halo mini PC: 44-59 tok/s with speculative decoding, ~1,400 tok/s prefill, engine is open(在新标签页中打开)
- Ugh I didn't want to post this... Back to Qwen3.8 27B(在新标签页中打开)
- Qwen3.8-Flash-Next on Strata(在新标签页中打开)
- 2x CMP 170HX 64GB: GLM-5.3-Flash at 384K context / ~90 tok/s (EXL3, HBM-first setup) + Qwen3.8 comparison(在新标签页中打开)
- $2800 rig with 8x Radeon Pro V620 (256 GB VRAM) + custom vLLM fork = Qwen3.8-Flash-Next at 60 to 100 t/s decode and 3000+ t/s prefill(在新标签页中打开)
- Qwen3.8-27B UD-IQ4_XS Heretic + MTP on a 16 GB card with 55-68 tok/s (24gb and 12gb versions available too)(在新标签页中打开)
- Qwen 3.8 Flash Next-GSQ-RCO-IQ2_XS at ~21 tok/s on just an RTX 3060 12GB + 16GB DDR4 RAM(No gate pruning, 100% bit-exact)(在新标签页中打开)
- Qwen3.8-Flash-Next-GSQ-RCO (IQ3_S): ~20-30 tok/sec decode & 300-90k tok/sec prefill on 12GB VRAM + 32GB RAM + NVME(在新标签页中打开)
- Qwen3.8 27b with 200K ctx + MTP on 12GB Ampere Cards(在新标签页中打开)
- Engineer / developer observations of Gemma4-31B, Qwen3.8-27B, and 6.1-Sol for software engineering work(在新标签页中打开)
- Benefits of using bigger models than Qwen 3.8 flash next?(在新标签页中打开)
- [vllm] Is it possible to run Qwen3.8-27B-FP8 on a 5090?(在新标签页中打开)
- 48Gb VRAM speed AND quality ! (Qwen 3.8 27B Swift 1.5 W8A16)(在新标签页中打开)
- Is anyone running Qwen3.8 Flash Next with a 1M context?(在新标签页中打开)
- Qwen3.8 Flash Next fixed my GNOME extension(在新标签页中打开)
- Qwen3.8-27B on a single 3090: 140 tok/s on code with a custom megakernel(在新标签页中打开)
AI 算力融资. Lambda 正以 145 亿美元的投前估值融资至多 40 亿美元,并计划在 2027 年 IPO。其积压订单在三个月内从 150 亿美元涨到 500 亿美元,主要来自 Anthropic 的 350 亿美元承诺。OpenAI 的年化收入正逼近 500 亿美元,低于此前的预期。
这是本周回顾 – 下周日见。