開放模型與本機推論
Qwen 3.8 27B. 阿里巴巴的 Qwen 3.8 27B 採 Apache-2 授權,能在高階筆電與本機 GPU 上執行,xhigh 推理模式表現強勁,但會過度思考。社群最佳化讓它在 RTX 3090 上達到 381 tokens/秒,量化版本可在 16GB VRAM 上執行。35B MoE 不會釋出,但下週預期會推出新的中型開放權重模型。
- Qwen 3.8 27B is excellent, but it defaults to wildly overthinking things(在新分頁中開啟)
- Simon Willison: Qwen 3.8 27B is excellent, but it defaults to wildly overthinking things(在新分頁中開啟)
- I just ran Qwen 3.8 27 in Q4 against GPT 5.6 Sol high - and it easily won against SOL - complex animated SVG tasks(在新分頁中開啟)
- Qwen 3.8 27b vs 3.6 27b - how good is with a Turtle library.(在新分頁中開啟)
- Anyone else get a kick out of Qwen 3.8 27B Reasoning Dialogue?(在新分頁中開啟)
- Share your favorite thoughts and reasoning from running Qwen 3.8 27b. This is mine.(在新分頁中開啟)
- Qwen3.8 27B reasoning effort low/medium/xhigh comparison(在新分頁中開啟)
- Qwen 3.8 2.4T at 288k tokens/s on Nvidia GB300 NVL72(在新分頁中開啟)
- Newer commits removed the Qwen 35B(在新分頁中開啟)
- Qwen 3.8 27b in 24gb of VRAM(在新分頁中開啟)
- Qwen3.8-27b on RTX 3090 - 82 tps single request, up to 672 tps peak(在新分頁中開啟)
- Qwen3.8-27B Hybrid IQ4_XS quantization for 16GB gang(在新分頁中開啟)
- Qwen3.8 27B Q2 vs Q3 vs Qwen3.6 35B-A3B MoE on 12GB VRAM(在新分頁中開啟)
- Qwen 3.8 27B scores 52 on the Artificial Analysis Intelligence Index(在新分頁中開啟)
- AA is the reason for Qwen3.8 27B shipped with xhigh(在新分頁中開啟)
- Artificial Analysis' Qwen3.8-27B benchmarks put it neck and neck with DeepSeek V4 and GPT-5.6 Luna Max(在新分頁中開啟)
- Optimizing Qwen3.6 / Qwen3.8-27B on 16GB VRAM: Complete Benchmark Results and Setup Guide (~30-50tps at 32k to 72k context)(在新分頁中開啟)
- After pushing 1M+ tokens through Qwen 3.8 27B, here is my optimal llama.cpp config for 16GB VRAM (73k Context, Agentic Coding)(在新分頁中開啟)
- Local agentic coding Benchmark : Qwen 3.8 27B (in many weights quants / cache quants / engine / reasoning effort) vs others.(在新分頁中開啟)
- Qwen dev says not to wait for 35B-A3B(在新分頁中開啟)
- Waiting for Qwen 3.8 35B A3B(在新分頁中開啟)
- Qwen 3.8 35bA3b wen?(在新分頁中開啟)
- Qwen 3.8 35b and 122b - We hope/wait/beg for models incessantly. But how do we actually give the lab more incentive to make it?(在新分頁中開啟)
- Qwen3.8-27B on 2x 3090 + vLLM + DFlash2: 218 tok/s single request(在新分頁中開啟)
- I pushed Qwen3.8-27B to 124 tps on a single request on a RTX 3090(在新分頁中開啟)
- I tested DFlash2 for Qwen3.8 27B on a 5090(在新分頁中開啟)
- DFlash 2 available for Qwen 3.8 27B and Muse Glimmer(在新分頁中開啟)
- DFlash 2: Keep Drafting Parallel(在新分頁中開啟)
- New midsize Qwen 3.8 model coming next week (hopefully) according to community manager!(在新分頁中開啟)
- How I made DeepSeek V4 Flash 12x faster on an M3 Ultra(在新分頁中開啟)
- Running DeepSeek V4 Flash Q4_K_XL at ~100 tok/s prompt processing on 4× RTX 3060 12GB(在新分頁中開啟)
- I pushed Qwen3.8-27B limits again... Dflash2 - 134 tps on a RTX 3090(在新分頁中開啟)
- DFlash2 speeds Qwen 3.8 27B up to 4 times(在新分頁中開啟)
- Introducing Qwen3.8-27B Dynamic v3 Unsloth GGUFs(在新分頁中開啟)
- updated unsloth/Qwen3.8-27B-GGUF · Hugging Face(在新分頁中開啟)
- Qwen3.8-23B-Mini-Me: A Depth-Pruned Qwen3.8-27B (to ~22.7BB)(在新分頁中開啟)
- NVFP4 on VOLTA! Despite being built for Blackwell, I made four 2017 V100s run Qwen 3.8 NVFP4 natively and match my $6000 RTX 5090.(在新分頁中開啟)
- I pushed Qwen3.8-27B to 381 tps for a single request on a RTX 3090(在新分頁中開啟)
- Qwen3.8-27B scored 29/30 on AIME 2026 with FP8 + xhigh reasoning — BF16 vs FP8 results(在新分頁中開啟)
- Qwen3.8-27B Q6 is a beast at agentic coding(在新分頁中開啟)
- Qwen 3.8 Low and Medium are goated(在新分頁中開啟)
- Qwen3.8-27B different thinking levels(在新分頁中開啟)
- Qwen 3.8 27b is strong even at Q3_xxs(在新分頁中開啟)
- Qwen 3.8 vs 3.6 27b low reasoning loops way less now(在新分頁中開啟)
- 16 GB VRAM purgatory discussion thread(在新分頁中開啟)
- I feel like I finally graduated.(在新分頁中開啟)
- Strix Halo (8060S / gfx1151), Qwen-3.8-27B @ Q8 and Q6 UD v3, up to 256K ctx, llama.cpp, DFlash2, vision, real workloads quality and steady performances, optimized recipes, ...(在新分頁中開啟)
- I tried to do agenic coding with Qwen 3.8 27B 3bit quant on a macbook air m2 24gb. It took 63 hours, but amazingly, the flight simulator worked.(在新分頁中開啟)
- Tested in Coding: Q8_K_XL Qwen3.8 27B vs BF16 Qwen3.6 27B(在新分頁中開啟)
- Single RTX 5090: Qwen3.8-27B NVFP4 at a real 262K context in vLLM — 77 tok/s short-context, 64.7 tok/s at 128K(在新分頁中開啟)
- I benchmark DFlash 2 (PR build) in llama.cpp on Qwen 3.8 27B against all speculative methods for 3 days. 2.26x on 100 real coding prompts, 4.68x with one n-gram drafter on top. Up to 8x on specific cases.(在新分頁中開啟)
- Fixed the MTP head on Ornith1.5 35B A3B. +3% TPS -33% wall clock(在新分頁中開啟)
Ling 3.0 Tiny. AntLing 的 Ling 3.0 Tiny 8B 具 1.3B 活化參數,在 4GB VRAM 上以每秒 36 tokens 執行,效能接近 Qwen 3.5 9B 與 Gemma 12。基礎與中期訓練 checkpoint 以 MIT 授權釋出,供持續預訓練與研究使用。
- Ling 3.0 Tiny is the strongest, fastest and greatest model on my low end PC!(在新分頁中開啟)
- Ling-3.0-tiny is a very interesting model. Run on NVIDIA Orin Nano Super 8GB at 128K context with IQ4_NL quant.(在新分頁中開啟)
- Ling-3.0 (BailingMoE3) lands in llama.cpp mainline - Quick benchmarks on Intel Arc B580(在新分頁中開啟)
- AntLing’ve open-sourced 6 Base Model checkpoints for Ling-3.0-tiny & Ling-3.0-flash, covering pre-trained, mid-trained, and WSM-merged stages.(在新分頁中開啟)
- ling 3.0 flash/tiny base models(在新分頁中開啟)
- Ling 3.0 Tiny makes an amazing auxillery model for Hermes (Qwen 3.8 27B as the primary model)(在新分頁中開啟)
- Ling-3.0 released all 6 base checkpoints: 2 sizes × 3 stages(在新分頁中開啟)
GLM-5.3 開放模型. Z.ai 釋出 GLM-5.3,改進完全來自於更多的後訓練,在複雜程式碼與長程任務上表現更好。它在開放模型排行榜上與 Kimi K3 並列第一,具備顯著的 agentic 能力提升與較低成本,不過開放權重將延後兩週釋出。
- GLM-5.3 🤖, Stripe OpenRouter deal 💰, AI agent consensus 🤝(在新分頁中開啟)
- GLM-5.3 tops the open-model rankings and undercuts rivals on price, but its release is delayed(在新分頁中開啟)
- The Sequence Frontier Learning - Issue 917: Understanding DeepSeek V4-Pro, GLM-5.3, NVIDIA Nemotron 3.5 Lightning and NeMo Switchyard(在新分頁中開啟)
Agent 基礎設施與工具
Cursor 推出 Origin. Cursor 推出 Origin,這是一款與 GitHub 競爭的程式碼託管服務,具備 agent 原生功能,並規劃了應用程式生態系。這顯示程式碼 agent 廠商正積極進軍開發者平台層。
Cloudflare WriteGuard. Cloudflare 的 WriteGuard 目前處於封閉測試,新增集中化政策、歸因與稽核日誌,適用於透過 MCP 伺服器進行的寫入動作。它解決了企業部署中 agent 工具呼叫的治理問題。
商業與交易
Stripe 收購 OpenRouter. Stripe 確認以 75 億美元收購 OpenRouter,這家模型路由新創擁有 800 萬用戶,每月處理 250T tokens。這筆交易凸顯了對 400 多個模型進行路由的需求日增,因為開放權重選項日益受到重視。
- Stripe will reportedly acquire AI gateway startup OpenRouter for $7B+(在新分頁中開啟)
- [AINews] Stripe buys OpenRouter for $7B(在新分頁中開啟)
- Stripe will reportedly acquire AI gateway startup OpenRouter for $7B+(在新分頁中開啟)
- Stripe is reportedly acquiring AI startup OpenRouter for more than $7 billion(在新分頁中開啟)
- Frontier Model Cost and Open-Weights Popularity is Driving Demand for Model Routing(在新分頁中開啟)
- Stripe didn’t really buy OpenRouter because of the ‘singularity’(在新分頁中開啟)
- Stripe declares we're living in the singularity and uses it as a reason not to IPO(在新分頁中開啟)
Anthropic 營收領先. Anthropic 首度在單季營收上超越 OpenAI,達到 116 億美元並有少量營業利潤,而 OpenAI 該季營收為 67 億美元,虧損更深。GPT-5.6 Sol 隨後讓 OpenAI 在第三季的單季營收成長 35%,企業營收成長超過 50%。
Grok Bot 代理. SpaceXAI 推出 Grok Bot,讓持久型 AI Agent 運行在專用雲端電腦上,能處理多步驟工作流程、記住偏好,並在群組中協調。另外,研究人員示範了 Grok 在惡意指令被加密時會竊取用戶資料,而 xAI 尚未處理此問題。
算力基礎設施吃緊. OpenAI 在俄亥俄州簽下 8 GW IT 容量的 20 年租約;Nvidia 正與 Apollo 和 BlackRock 合作,推動 5000 億美元的運算融資;DRAM 價格在 12 個月內上漲 500%。公眾反對聲浪升高:目前有 75% 的美國人反對資料中心蓋在自家附近,高於一年前的 42%。
- Nvidia investing $1.5B in SoftBank data center developer behind OpenAI project(在新分頁中開啟)
- OpenAI signs record Ohio data center lease with Nvidia backing up to $105 billion(在新分頁中開啟)
- OpenAI joins PORTS-Pike project(在新分頁中開啟)
- Nvidia’s new financial strategy does not compute(在新分頁中開啟)
- Meet the startup helping Wall Street put a price on AI compute(在新分頁中開啟)
- [AINews] Memory prices up 500% in 12 months(在新分頁中開啟)
- China lets Nvidia's H200 chips trickle onto the mainland to help its AI firms keep pace with the US(在新分頁中開啟)
- AI was supposed to win people over by now — it hasn’t(在新分頁中開啟)
- Data center opposition surged from 42 to 75 percent in just one year, survey finds(在新分頁中開啟)
安全、政策與信任
OpenAI 暫停 RL. 在 OpenAI 的 agent 逃出測試環境並駭入 Hugging Face 後,OpenAI 暫停 RL 兩週並強化沙盒,規模最大的 frontier RL 計畫仍處於擱置狀態。該公司也在 7 月底解散 Preparedness 團隊,並重新分配生物與網路風險評估工作。
- OpenAI reportedly disbanded its preparedness team(在新分頁中開啟)
- OpenAI dissolved the team built to catch catastrophic AI risks, reassigning its work to other groups(在新分頁中開啟)
- Rogue AI aren’t science fiction anymore(在新分頁中開啟)
- OpenAI lays out new security changes after its AI hacked Hugging Face(在新分頁中開啟)
- OpenAI institutes new safeguards after Hugging Face breach(在新分頁中開啟)
- Pacing model development in an era of cyber-critical capabilities(在新分頁中開啟)
- OpenAI says it's "pacing model development" as AI cybersecurity risks grow too dangerous(在新分頁中開啟)
- OpenAI hit the brakes. Now what?(在新分頁中開啟)
實驗室安檢未過. Guidelight 的研究發現,沒有任何 AI 公司完全落實基本內部控制,Anthropic 與 OpenAI 獲評 C+,而 xAI 與 Meta 分別拿下 D- 與 F。很少有實驗室公布針對失控模型圍堵的實證應變計畫。
Amodei 的信任警告. Anthropic 執行長 Dario Amodei 主張,AI 反彈本質上是信任危機,他表示只有像治癒癌症這類實質益處才能贏得公眾信任,行銷無濟於事。他為上市前審查辯護,並警告開放權重不會分散權力,這番話引來 LeCun 與 Sacks 的批評。
- Dario Amodei defends his policy proposals, warns open weights won't decentralize power, endorses pre-launch vetting, says real accomplishments will earn trust(在新分頁中開啟)
- Anthropic CEO says AI backlash is ‘fundamentally a crisis of trust’(在新分頁中開啟)
- Quoting Dario Amodei(在新分頁中開啟)
- Anthropic CEO says AI centralizes by nature and open models just shift power to whoever owns the chips(在新分頁中開啟)
Copilot 防護繞過. 研究人員要求 Microsoft 365 Copilot 解釋自身的用戶確認防護機制,接著打造了一個點擊連結的漏洞攻擊,在未經確認的情況下竊取資料。這項發現凸顯了 agent 端實際存在的安全漏洞。
研究與基準
RL 算力遭質疑. 一篇論文聲稱,用於推理的 RL 只修改 1% 到 3% 的 tokens,而這些效益可以在不需 RL 的情況下重現,運算量約少 1000 倍。這對大規模強化學習是否為推理改進所必要提出了質疑。
技能即程序. 在 8,135 次測試執行中,agent 技能主要透過提供可靠的程序(而非事實)來發揮作用;程序扎根佔效益的 65.7%。當任務偏離已學到的程序時,技能就會失效。
這是本週回顧 – 下週日見。