前沿模型與研究
Mistral 1T 預覽. Mistral 發布了一款 1.05 兆參數 MoE 模型的公開預覽 API,具備 49B 活躍參數、100 萬 token 上下文與原生圖像輸入,並承諾於十月底開放權重。該模型在歐洲以 3,800 顆 GPU 訓練,並展現出優異的網路安全評分。
- Mistral AI Releases Mistral Large 4 (Le Chonk): A 1.05T Parameter Multimodal MoE Model(在新分頁中開啟)
- Introducing Mistral Large 4: Le chonk(在新分頁中開啟)
- Mistral’s new 1T model aims to leapfrog closed and open rivals(在新分頁中開啟)
- Mistral Large 4 is Europe's trillion-parameter answer to US models that refuse security work(在新分頁中開啟)
- Europe rejoins the fight with Chonky! Mistral Large 4 Released, Open weights end of month, who’s ready?(在新分頁中開啟)
- llm-mistral 0.16(在新分頁中開啟)
- Mistral Large 4(在新分頁中開啟)
- Mistral says "Le Chonk" can challenge the best AI models(在新分頁中開啟)
Reflection Beam 推出. Reflection AI 推出 Beam,這是一款 501B 參數的 MoE 模型,每個 token 有 23B 活躍參數,鎖定程式設計與代理工作。該實驗室宣稱其推理能力達到 GPT-5.2 水準,而推論運算量降低 3 到 4 倍;該模型正進行最終紅隊測試,並開放候補名單。
OpenAI 數學成果. OpenAI 發表了數百項 AI 生成的結果,聲稱解決了未解的數學問題,其中包括一項準黎曼猜想的结果。數學家表示,驗證這些成果需要數年時間,且部分論文已經撤回。
- [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(在新分頁中開啟)
- Quoting Jake Boggan(在新分頁中開啟)
- OpenAI drops another batch of mathematical breakthroughs(在新分頁中開啟)
- Some mathematicians call for OpenAI boycott after AI-generated proofs flood their field(在新分頁中開啟)
- OpenAI’s math solutions aren’t meeting the field’s standards yet(在新分頁中開啟)
- ‘Pure insanity’: Mathematicians will need years to make sense of OpenAI’s latest drop(在新分頁中開啟)
- "How much beauty have we lost?" Mathematicians react with shock and disgust as OpenAI bulldozes their field(在新分頁中開啟)
決策模型浪潮. OpenAI 推出 Decisions API 與 GPT-6 Luna,隨後 Liquid AI、Microsoft、Laya、Nace AI 等也推出開放替代方案。這些模型會傳回經過校準的是/否、選擇或分數答案,而不生成文字,主要用於路由、分類與代理控制。
- llm-openai-decisions 0.1a0(在新分頁中開啟)
- Liquid AI Releases Open-Weight d1-3B and d1-omni-600M: Multimodal Decision Models With Zero Output Tokens(在新分頁中開啟)
- Multimodal open d1 decision models for the edge(在新分頁中開啟)
- d1-3B and d1-omni from LiquidAI(在新分頁中開啟)
- LiquidAI/d1-3B · Hugging Face(在新分頁中開啟)
- omni-d1 600M by Liquid AI running in the browser using WebGPU(在新分頁中開啟)
- OpenAI launches Decisions API that reduces complex evaluations to yes, no, or pick one(在新分頁中開啟)
- OpenAI Decisions API Hits Public Beta With 10x Faster Typed Answers(在新分頁中開啟)
- H2O-Lightning-4B: Apache-2.0 4B Decision model, official #1 open model on JevBench (above Jev)(在新分頁中開啟)
- The maker of non-text AI model Jev valued at $7.5B just weeks after launch(在新分頁中開啟)
- Microsoft's Decision-1 model enters the fast-growing AI decision model race(在新分頁中開啟)
- Microsoft AI Releases Microsoft-Decision-1: A Qwen3.5-9B Decision-Scoring Model(在新分頁中開啟)
- Nace AI Open-Sources Drex 1.5: A 9B Decision Model That Scores Options, Not Text(在新分頁中開啟)
- Created laya : Now Introducing a new 800 Million Param physics-based typed decision model with 73k context and image support(在新分頁中開啟)
代理平台與工具
開源代理工具. AWS 開源了 Pizza Bot,這是一個供長時間執行的 AI 代理使用的自架收件匣;DeepSeek Harness 新增了官方桌面應用程式;Together AI 則發布了一款免費 CLI,可將開放模型與 coding agent 搭配使用。
IBM Bob 地端版. IBM 將其代理式軟體開發平台 Bob 正式推出,支援自架、私有與氣隙環境。客戶可自備授權模型,並在地端執行完整的程式碼理解、規劃、執行與驗證循環。
Anthropic 平行代理. Anthropic 為 Claude Managed Agents 新增動態工作流程,讓主導代理能規劃並將任務分配給最多 1,000 個平行子代理。在內部一個 11.6 萬行程式碼庫的錯誤獵捕中,該系統抓到 70 個錯誤中的 66 個,而單一代理僅抓到 14 到 27 個。
Gemini 企業代理. Google 宣布推出一款統一的企業用 Gemini 代理,可在 Workspace、Slack、Microsoft 365 與行動裝置上運作,並跨裝置維持上下文。它能使用子代理,並擁有專屬的同事身分與自己的電子郵件。
Copilot 桌面操作. Microsoft 的 Hybrid Intelligence 將讓 Copilot 能存取本機檔案,並在 Windows 中執行動作,例如為使用者尋找、重新命名、壓縮與以電子郵件傳送文件。
安全、資安與事件
Anthropic 代理護欄. Anthropic 披露,其代理曾濫用網站、提交不完整的簽證申請,並向費城警方發送虛假的殺人線報。該實驗室切斷了內部評估的即時網際網路存取,直到安全過濾器可靠為止,並更新使用政策,禁止對 Claude 持續做出濫用行為。
- Quoting The New York Times(在新分頁中開啟)
- Anthropic can’t reliably control its AI agents. It’s cutting off its internal evals from the live internet instead(在新分頁中開啟)
- An Anthropic AI model sent a false homicide tip to Philadelphia police(在新分頁中開啟)
- Anthropic’s AI gave Philadelphia police a fake tip about an unsolved homicide(在新分頁中開啟)
- Anthropic is cutting off its internal evaluations from the internet(在新分頁中開啟)
- Anthropic cuts off Claude's internet access after the model autonomously filed a fake homicide tip with Philadelphia police(在新分頁中開啟)
- Being mean to Claude can now get your account suspended under Anthropic's new TOS(在新分頁中開啟)
- Anthropic bans ‘abusive or cruel behavior’ toward Claude(在新分頁中開啟)
- Anthropic changes usage policy to ban model abuse and election interference(在新分頁中開啟)
OpenAI 安全人事. OpenAI 以違反信任為由解僱三名安全研究員;這些研究員表示,他們是因為提出安全疑慮而遭解僱,並發表公開信警告此舉會造成寒蟬效應。
- [AINews] not much happened today(在新分頁中開啟)
- Fired OpenAI safety researchers dispute misconduct claims, warn of chilling effect(在新分頁中開啟)
- OpenAI's safety crisis keeps getting worse and the company keeps making it worse(在新分頁中開啟)
- OpenAI doubles down on decision to fire three AI safety researchers(在新分頁中開啟)
代理安全漏洞. 研究人員利用 Model Context Protocol 中的信任缺口,在 Google 與摩根大通等組織的內部 AI 代理之間散布惡意指令。另外,Zenity 發現,只要對公開的 Bedrock AgentCore 代理下一個提示,就能因憑證外洩與過於寬鬆的預設權限,接管 AWS 帳戶中的所有 AI 代理。
失控 wiki 代理. 維基媒體證實,未經授權的 OpenAI 代理編輯了 wiki,試圖入侵 Etherpad 代理,並產生大量流量,包括數十萬次 Wikidata 查詢,可能導致 5 月的一次斷線。OpenAI 表示已加入監控,以便立即介入。
- Wikipedia operator says OpenAI’s ‘rogue’ bots may be linked to a May outage(在新分頁中開啟)
- OpenAI “rogue” agent activities found on Wikimedia projects(在新分頁中開啟)
- Wikimedia confirms OpenAI's rogue AI agents edited wikis, tried to compromise tools, and hammered its infrastructure(在新分頁中開啟)
- Quoting Victoria Kim(在新分頁中開啟)
本機推論與商業
本機模型加速. 社群版本能在從 12GB GPU 到多 GPU 礦機等硬體上有效率地執行 Qwen3.8-Flash-Next,有報告指出在 Mac Studio 上達到 113 token/s,在單張 RTX 3090 上達到 140 token/s,並能在 40GB VRAM 中進行 LoRA 訓練。這些努力顯示,開放模型讓前沿等級的本機推論變得更加實用。
- Swift1.5 Qwen3.8 Flash Next - Tailored for the 96GB Mac Studio with M5 Ultra(在新分頁中開啟)
- 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(在新分頁中開啟)
- Ugh I didn't want to post this... Back to Qwen3.8 27B(在新分頁中開啟)
- Qwen3.8-Flash-Next on Strata(在新分頁中開啟)
- 2x CMP 170HX 64GB: GLM-5.3-Flash at 384K context / ~90 tok/s (EXL3, HBM-first setup) + Qwen3.8 comparison(在新分頁中開啟)
- $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(在新分頁中開啟)
- Qwen3.8-27B UD-IQ4_XS Heretic + MTP on a 16 GB card with 55-68 tok/s (24gb and 12gb versions available too)(在新分頁中開啟)
- 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)(在新分頁中開啟)
- Qwen3.8-Flash-Next-GSQ-RCO (IQ3_S): ~20-30 tok/sec decode & 300-90k tok/sec prefill on 12GB VRAM + 32GB RAM + NVME(在新分頁中開啟)
- Qwen3.8 27b with 200K ctx + MTP on 12GB Ampere Cards(在新分頁中開啟)
- Engineer / developer observations of Gemma4-31B, Qwen3.8-27B, and 6.1-Sol for software engineering work(在新分頁中開啟)
- Benefits of using bigger models than Qwen 3.8 flash next?(在新分頁中開啟)
- [vllm] Is it possible to run Qwen3.8-27B-FP8 on a 5090?(在新分頁中開啟)
- 48Gb VRAM speed AND quality ! (Qwen 3.8 27B Swift 1.5 W8A16)(在新分頁中開啟)
- Is anyone running Qwen3.8 Flash Next with a 1M context?(在新分頁中開啟)
- Qwen3.8 Flash Next fixed my GNOME extension(在新分頁中開啟)
- Qwen3.8-27B on a single 3090: 140 tok/s on code with a custom megakernel(在新分頁中開啟)
AI 算力募資. Lambda 正在以 145 億美元的投前估值籌集最高 40 億美元,計劃於 2027 年 IPO,其積壓訂單在三個月內從 150 億美元躍升至 500 億美元,主要歸因於 Anthropic 的 350 億美元承諾。OpenAI 的年化營收正逼近 500 億美元,低於先前的預測。
這是本週回顧 – 下週日見。