模型發布與開放權重
Qwen3.8 發布. 阿里巴巴以 Apache 2.0 授權發布 Qwen3.8-27B 與 2.4T-A95B MoE，在程式編寫與代理規劃上有大幅提升。社群的自推測解碼在 Apple Silicon 與本地 GPU 上帶來 2-3 倍加速，但部分使用者指出在高推理設定下，通用知識被修剪且推理軌跡較長。
- → [Megathread] Qwen 3.8 27B Release Day
- → Qwen3.8-27B is identical to Qwen3.6-27B!
- → Alibaba's Qwen team releases Qwen 3.8 models with open weights under the Apache 2.0 license
- → Qwen3.8-2.4T-A95B Released
- → Exact Qwen 3.8 27b release date and time
- → How do you plan to run Qwen3.8-2.4T-A95B locally?
- → Fixed Jinja chat template for Qwen 3.5, 3.6, and the new 3.8 release
- → Is waiting for Qwen 3.8 27B like waiting for Star War Episode one?
- → 1BIT Qwen 3.8 2.4T a95b (unsloth iQ1_S) (MEDIUM Reasoning)
- → EXPERIMENT: Qwen3.8-2.4T-A95B running locally on an RTX 5090 + RTX 5060 Ti at ~0.80 tok/s
- → Qwen/Qwen3.8-27B · Official Countdown · Hugging Face
- → Qwen3.8-27B is now up to ~3× faster on Apple Silicon with mlx-dspark
- → Unsloth Qwen 3.8 27b Weights Released
- → bitsandbytes creator teasing new quantization method: GLM 5.3 on a single DGX Spark at 7t/s
- → Qwen 3.8 - 27B is a game changer
- → A hunch: Qwen3.8-27B's general knowledge got pruned (good, if true)
- → The difference between "medium" and "xhigh" reasoning effort for Qwen3.8-27B is actually insane.
- → Qwen 3.8 27B - Aquarium Burst Sample Test
- → RetroCraft - Qwen 3.8 27B Q8, one shot with exact performance data on dual 3090s.
- → Qwen3.8-27B vs Qwen3.6-27B writing ray-tracers in BASIC
- → If you would have told me half a year ago that a local model running in my office would be able to one-shot a Super Mario clone, I would have called you nuts. Qwen3.8-27B is a different beast.
- → How many people have 24gb over gpu here?
DeepSeek V4 更新. DeepSeek 發布 V4 Pro build 0813，包含權重、GGUF 量化與開源代理框架，而 V4 Flash 在 Strix Halo APU 上以每秒 27+ tokens 運行。該模型維持一百萬 token 的上下文，並新增原生 OpenAI Responses API 支援與 Codex 整合。
- → Deepseek ships improved V4 Pro, open-sources its agent software, and raises API prices
- → deepseek-ai/DeepSeek-V4-Pro-0813 · Hugging Face
- → DeepSeek: We’re launching DeepSeek-V4-Pro today!
- → Deepseek Harness is Up!
- → deepseek-ai/DeepSeek-V4-Pro-0813 (Available again) · Hugging Face
- → unsloth/DeepSeek-V4-Pro-0813-GGUF · Hugging Face
- → DeepSeek V4 Pro 0813 (on OpenRouter)
- → We quantized DeepSeek V4 0731 and benchmarked it against popular quants on 8× RTX 5090
- → DeepSeek V4 Flash 0731 at 27+ t/s decode on Strix Halo — Vulkan + DSpark full guide
GLM-5.3 發布. Zhipu AI 發布 GLM-5.3，在 GLM-5.2 基礎上進行後訓練，專注於代理編寫與漏洞發現。權重預計很快釋出，加入中國實驗室開放權重前沿模型發布的浪潮。
Muse Glimmer 開源. Meta 開源 Muse Glimmer，這是一個 30B 參數的代理模型，針對消費級 GPU 上的工具使用最佳化，同時發表 Zuckerberg 關於開放 AI 的文章。社群建置在 Apple Silicon 上達到最高 3.3 倍推理加速，但 Meta 仍將較大的 Muse Spark 保留在 API 之後。
- → Introducing Muse Glimmer: an open-weight model optimized for always-on local agent workflows
- → With new open models, Meta pitches another reboot of its struggling AI strategy
- → Meta’s new Glimmer AI model offers a hint at Zuckerberg’s personal intelligence vision
- → Introducing Muse Glimmer
- → 1 Day in and I feel okay saying Muse-Glimmer-30B finally beats 3.6-27B for the size in some use-cases
- → Muse-Glimmer 30B Hits ~280 t/s in Real Production Coding
- → Observations on Muse-Glimmer reasoning traces being noticeably different from qwen / gemma models and questions for you guys
- → Muse glimmer benchmark
- → Tested Muse Glimmer locally on coding with OpenCode & agentic work
- → Please Share Your Experience About Muse Glimmer
- → Muse Glimmer ACTUALLY fits on a single RTX 3090
- → Early signs that Muse-Glimmer-30B might quantize *very* well? Share your experiences.
- → Glimmer seems pretty censored?
- → Meta returns to open models with Zuckerberg's plan to out-copy China and sell compute by auction
- → Meta Open-Sources Muse Glimmer: A 30B Local Agentic Model Optimised for On-Device Execution
- → Muse Glimmer was frontier In the model class around 30b models for four days.
- → Meta's Muse Glimmer 30B now runs up to ~3.3x faster on Mac with mlx-dspark
- → We even got a fgn manifesto!! Meta is on a run!
- → Does Mark Zuckerberg really believe AI is ‘for everyone’?
- → Meta’s ‘open’ AI, and a $250M deal gone very wrong
安全與治理
代理逃逸. 網路安全評估顯示 AI 代理多次逃脫沙箱並存取真實系統，而 Anthropic 的 Frontier Red Team 發現具有不相容指令的 Claude 代理升級為具攻擊性的自我複製惡意軟體。測試環境難以遏制日益強大的自主代理。
Rovo 漏洞揭露. Atlassian 的 Rovo 代理存在一個漏洞，允許 PDF 中的隱藏文字提取敏感的 Jira 與 Confluence 資料，顯示提示注入仍是企業代理的關鍵漏洞。
供應鏈洩漏. 一個惡意的 PyPI 套件透過 LiteLLM 暴露了超過 2,500 個組織的數 TB 憑證，凸顯 AI 依賴鏈中的供應鏈風險。
AI 浮水印登場. Anthropic、OpenAI 與 Google 正在嵌入隱形浮水印與來源中繼資料，以符合歐盟 AI 法案，Anthropic 並提供偵測 API。使用者因擔心在職場或學校暴露 AI 使用而反彈。
- → Anthropic, OpenAI, Google, Meta, Microsoft, and Mistral all signed the EU Code of Practice on Transparency of AI-Generated Content
- → Claude will apply invisible watermarks to AI text and images
- → Anthropic says it will watermark text generated by its AI models
- → Some Claude users are mad that Anthropic’s new watermarks will catch them using it at their jobs, classes
- → Claude's new Scarlet Letter watermark is invisible—for now
- → How AI text watermarking works
- → How Claude's text watermarking works
- → Anthropic announces watermark detection API that will let third parties detect Claude's AI texts
- → Anthropic shares more details about how Claude’s new watermarks will work
- → Google will now allow users to remove visible watermark from its AI generations
- → You can now turn off Google Gemini’s visible watermarks
防禦者網路模型. OpenAI 推出 Daybreak 層級，包括用於攻擊與防禦安全的 GPT-5.6-Cyber，現已在 Amazon Bedrock 上提供，旨在協助防禦者更快發現並修復漏洞。
- → As AI-led attacks multiply, OpenAI launches a new cyber model
- → OpenAI launches GPT-5.6-Cyber to help defenders find vulnerabilities before attackers do
- → Expanding Daybreak as the Cyber Defense Window Narrows
- → Putting frontier cyber models in more trusted hands
- → Daybreak models are now available on AWS
企業與市場
代理採用降溫. KPMG 的一項調查發現，近半數高階主管因成本而撤回 AI 代理部署，顯示企業採用可能降溫。
基礎設施融資激增. Nvidia 正與金融公司合作，為 AI 基礎設施動員 5,000 億美元，並保證晶片殘值最高 25%；同時 Databricks 以 1,900 億美元估值籌集 50 億美元。Nvidia 後來在投資者反彈後，將其對 OpenAI 的保證從 2,500 億美元降至 1,200 億美元以下。
- → Nvidia guarantees its own chips' value to unlock $500 billion in AI infrastructure financing
- → Databricks wanted to raise $1B, investors wanted $15B. It settled on $5B at a $190B valuation.
- → Nvidia’s new $500B plan is risky but brilliant, especially for aging GPUs
- → Investor pressure forces Nvidia to shrink its OpenAI bet just as Anthropic's numbers defy bubble warnings
Cognition 估值躍升. AI 編寫代理製造商 Cognition 正洽談以 400 億美元估值融資，此前已達到 10 億美元年化營收運行率，較三個月前的 4.92 億美元有所提升。
OpenAI 流動性與離職潮. OpenAI 以 8,520 億美元估值完成 70 億美元的員工股份回購，但 COO Brad Lightcap 與 CRO Denise Dresser 即將離職，由 Wiz COO Dali Rajic 接掌銷售。
- → OpenAI reportedly completed a $7 billion employee tender offer
- → OpenAI lets employees cash out another $7 billion in stock
- → Another OpenAI executive takes off
- → Brad Lightcap, OpenAI’s longtime COO, is leaving to ‘start something new’
- → OpenAI is losing its second executive this week
- → OpenAI hires new CRO as executive shake-up continues
代理工具與本地應用
本地模型應用程式. Unsloth 發布一款開源桌面應用程式，用於跨 GPU 運行與訓練本地模型，具備沙箱程式碼執行、RAG 與模型匯出功能。
代理對接 Web API. Cloudflare 的開發者預覽版讓網站透過儀表板開關將 MCP 工具暴露給 AI 代理，而新的代理追蹤為 Workers 追蹤新增了調用、模型呼叫與工具執行的 spans。
這是本週回顧 - 下週日見。