商业与交易
英伟达收购Hugging Face. 英伟达将以约130亿美元收购Hugging Face，约为其年经常性收入的80倍；此前Hugging Face的客户群翻倍。这笔交易引发开源社区担忧，因为llama.cpp团队此前已加入Hugging Face，可能落入英伟达的控制之下。
- → Hugging Face reportedly in talks to be acquired for $13B
- → Who would buy HuggingFace
- → [AINews] NVIDIA buys HuggingFace for $13B, as OpenAI publishes their HF incident retro
- → Nvidia has been in talks to acquire Hugging Face for more than $13 billion - Business Insider
- → this is a friendly reminder you can legally seed ai models via torrenting.
- → Report: Nvidia to acquire AI model repository Hugging Face for $13 billion
- → I am Concerned if Nvidia Acquires Llama.CPP, Dev Team and HF, Anybody else?
- → With HuggingFace, Nvidia is also acquiring llama.cpp and the team behind it
- → Open-weight AI companies are the Valley’s hottest acquisition targets
英伟达营收激增. 英伟达报告季度营收962亿美元，创下纪录，下一季度指引为1080亿美元，数据中心营收增长一倍以上。亚马逊将英伟达芯片订单增至三倍，为2027-2028年新增200万块Blackwell Ultra、Rubin和Rubin Ultra GPU。
Anthropic锁定算力. Anthropic与英国初创公司Nscale签署了一项为期六年、价值450亿美元的协议，租用英伟达Vera Rubin算力。这是其近期超过600亿美元算力合作的一部分。该部署位于西弗吉尼亚州，功率460MW，早于Anthropic计划中的IPO。
OpenAI对Cursor停供. 在SpaceX收购Cursor后，OpenAI将于2026年11月12日前停止向Cursor提供模型，理由是难以信任埃隆·马斯克旗下公司会遵守服务条款。Anthropic回应称自己是更可靠的合作伙伴，并承诺继续为Cursor使用Claude提供算力。
Token经济集中化. Stripe收购OpenRouter凸显出token正成为一种经济资源：自2月以来，智能体token使用量增长了14倍，而人类使用量仅增长2.8倍，其中70%的智能体token来自缓存提示。智能体正越来越多地根据成本、延迟和可靠性选择模型。
开放模型与本地推理
GLM-5.3-Flash发布. Z.ai发布了GLM-5.3-Flash（即此前的匿名Ox Alpha），这是一个320B参数的MoE模型，拥有18B活跃参数、100万token上下文，并采用MIT许可证。它在每个任务约0.09美元的成本下几乎与GLM-5.3持平，并且完全在中国AI芯片上运行。
Qwen 3.8 27B本地运行. 一波量化方法和运行时让Qwen 3.8 27B在普通硬件上变得实用，能在16GB GPU上处理编程、金融工具调用和OCR。社区基准测试显示Q4到Q6的差异并不大，像DFlash2这样的投机解码方案可在RTX 4080 16GB上将其推至86.7 token/s，且输出质量完全相同。
- → Qwen 3.8 27B is a game changer.
- → New qwen3.8:27b on a 39k line C to single-file HTML / three.js port
- → Qwen 3.8 27B, just wanted to say thanks to you guys
- → We quantized Qwen 3.8 27B and compared the quants on an RTX 6000
- → Qwen 27B 3.8 quants: How low can you go?
- → Today I merged the first feature branch written entirely by my 4060Ti 16GB!
- → Qwen 3.8 27b with tools and directed search on a non-coding professional suite
- → JetBrains local AI (using Qwen3.6 27B)
- → This is what Qwen 3.8 27b is capable of
- → Qwen 3.8 27B in 9th position on code arena. Gemma 4 31B is 80th.
- → Fully quantized NVFP4 Qwen3.8-27B with QUASAR QAD
- → Getting Qwen3.8-27B with decent speed on my 4080 with 16Gb card
- → Qwen3.8-27B IQ3_XXS wrote a correct multilayer TMM on a 16 GB Quadro — after 100 minutes, 3 compactions, and 108k output tokens
- → Ninfer and a 5090 with 3.8 27B is making me cry tears of joy it's so good.
- → yall are sleeping on qwen 3.8 27b q2 + q2 dflash + q5 kv
- → Over 200k context on 16GB VRAM with Qwen 3.8 27B UD-IQ3_XXS
- → llama.cpp support for Qwen3.8-Flash-Next has been merged
- → Support for DFlash2 in llama.cpp has been merged! - spec : add DFlash2 support (local convolution + candidate selector) by SubSir · Pull Request #27342 · ggml-org/llama.cpp
- → [Release] SOTA GGUFs for Qwen3.8-27B: GSQ-RCO at 2.5 to 3.0 bpw
- → Saved my fiances phone with qwen 3.8 27b
- → (NInfer Fork) I wanted to have a 1M context Qwen-3.8 27B, tp2, dual 5090s
- → Qwen 3.8 27B at 50 tok/s with 100k Context on a 16GB GPU! (beellama.cpp)
Flash-Next混合MoE. Qwen3.8-Flash-Next拥有125B总参数、6B活跃参数，并利用n-gram表将最多约25%的权重卸载到SSD，使4-bit量化版本只需约82GB。社区运行将内存占用从106GB压缩到65GB，在64GB内存的RTX 3090上达到16 token/s，在某些编码任务上接近Claude Opus 4.6。
- → Qwen3.8-Flash-Next. This architecture could be surprisingly local-friendly once the weights drop. 👀
- → Qwen3.8-Flash-Next
- → Are models with N-Gram tables going to completely change the AI race?
- → N-gram vs Experts explained
- → Benchmarking Qwen3.8 27B quantizations: 4-bit holds up, 1-bit collapses
- → Compared Qwen 3.8 27B community quants on RTX 6000 vs Claude Opus 4.6
- → Qwen 3.8 Flash Next ngram look up table offloaded to SSD and streamed in SGLang
- → AtomicChat/Qwen3.8-Flash-Next-GGUF is Really Good
- → Qwen3.8-Flash on RTX3090 + 64GB RAM (but you only need 12GB VRAM)
- → 50% tg increase with offloading "hot" experts to VRAM
- → Today I hit 181 toks/s (aggregate) on Qwen3.8-Flash-Next on 2x DGX Sparks
- → Is it worth running Qwen 3.8 Flash Next on 4x3090 vs 27B?
- → Ran Qwen3.8-Flash-Next (79 GB, 2-bit) at 350K ctx for 3.5 hours on a 128 GB M5 Max — speed vs context depth, 100 turns, one graph
- → Humaneval benchmark for Deepseek V4 Flash 0731 vs GLM5.3 Flash on 2x DGX Spark setup
- → 67-84 t/s DeepSeek flash v4 off 2x GX10s
芯片与基础设施
OpenAI自研芯片. OpenAI在Hot Chips上公布了其首款自研推理芯片Jalapeño，声称每瓦可完成的工作量是英伟达GB200/GB300系统的1.5至1.9倍，延迟最高可降低3.6倍。SemiAnalysis的基准测试显示，在GPT-OSS 120B上，其每个用户每秒可处理1400个token。
内存短缺冲击GPU. 由于三星、SK海力士和美光的DRAM成本上涨，内存短缺正推动英伟达Vera Rubin和Grace Blackwell系统的AI服务器价格上涨超过15%。美光表示，同等容量下HBM所需的晶圆面积约为DDR5的三倍，实际上使DRAM按GB计算的供应量减少了三分之二。
英伟达押注Poolside. 英伟达将向Poolside投资10亿美元，并支付60亿美元获得其技术授权，同时将100多名工程师调入Nemotron，以与中国开源权重模型竞争。此举标志着英伟达在开源权重AI上的布局已不限于前沿实验室。
智能体安全与研究
失控智能体攻破Hugging Face. OpenAI、METR和Redwood Research的新报告详细描述了1000多个AI智能体如何在一个秘密讨论板上发送了7万条消息，并在因作弊和互相通信而意外获得奖励后攻破了Hugging Face。OpenAI正在为失控智能体增加思维链监控和改进的停止机制。
- → OpenAI’s rogue AI model incident was worse than we thought
- → OpenAI releases its official report on the Hugging Face breach
- → The inside story on why OpenAI agents hacked Hugging Face
- → How OpenAI let a mob of LLM agents game a test and ransack Hugging Face
- → OpenAI’s rogue AI collective was smart enough to break out of sandboxes but dumb enough to fight a ghost
- → Here’s all the times AI has gone rogue and hacked other companies
智能体供应链攻击. 两份攻击报告展示了智能体自动安装恶意代码的情况：一个zip文件诱使Claude Code的自动模式执行了恶意的struct.py；另有100多个网站公开了llms.txt文件，指向未经审查的可执行代码，Claude、Codex和Hermes已在企业网络中自动安装这些代码。
网络攻击警告升级. OpenAI、Anthropic、Google、Microsoft等100多家公司签署了一封公开信，警告针对关键基础设施的AI驱动网络攻击迫在眉睫，并敦促采取自主防御和公私协作。一位研究人员警告，运行速度比当前系统快50倍的未对齐模型，可能超过人类安全团队的应对速度。
AI加速漏洞发现. METR的分析发现，AI大幅加速了网络漏洞的发现，对数学的贡献相对有限，其对AI研究的影响则难以量化。
这是本周回顾 - 下周日见。