AI Daily · Oct 30, 2026 · Friday
BAD INFLUENCE
vLLM adds day-0 support for Mistral Large 4 with DSpark and NVFP4
vLLM announces day-0 support for Mistral Large 4, a 1-trillion-parameter mixture-of-experts multimodal model with 49 billion active parameters per token. The post reports about 2x output throughput on TP8 with the built-in DSpark drafter and about 1.5x more KV cache with the NVFP4 checkpoint, which stays within 2.1 points of FP8 on evaluated tasks.
Model releases & updates
Mistral releases open weights for improved Mistral Large 4 model
Mistral AI has released the open weights of an improved Mistral Large 4, a 1-trillion-parameter multimodal model, after three weeks of additional reinforcement learning. The company reports a 51% score on the Artificial Analysis Cyber Index and a 41% score on Terminal-Bench 4.0, up 13 points from the public preview. The model is priced at $0.70 per million input tokens and $2.10 per million output tokens, and is available under a modified MIT license.
Product launches & updates
LangChain BlogOfficial3 more sourcesLangChain gives Managed Deep Agents a browser with handoff for sign-in
LangChain says Managed Deep Agents now lets agents use a browser for tasks such as booking reservations and end-to-end testing. Developers add one define_browser export, and a browser_handoff tool passes control to a person for sign-in. On 100 WebVoyager tasks, LangChain reports a 90% pass rate for $1.98 total, the highest of seven toolsets it tested.
Rohan Paul@rohanpaul_aiX2 more sourcesAnthropic trains Claude to flag its own potential misuse after each reply
Anthropic says it has trained Claude to give a preliminary safety assessment of each conversation right after responding. The assessment flags misuse that its other safeguards were not designed to catch, and the company says it is efficient because Claude has already read the conversation. The author's post adds that the self-assessment matches 78% of a costlier classifier's flags on Fable 5 traffic, rising to 94% for high-priority cases, at a 1% false-alarm rate.
Hugging Face releases AnySpeechRecognizer and AnyDecisionModel for Apple platforms
Hugging Face introduces two open Swift packages, AnySpeechRecognizer and AnyDecisionModel, for running speech recognition and yes-or-no, multiple-choice, and rating decisions on Apple devices. AnySpeechRecognizer keeps Apple's SpeechAnalyzer API and lets developers choose NVIDIA Parakeet or Alibaba Qwen3-ASR as the engine, with speaker diarization added. AnyDecisionModel derives typed answers and probabilities from model outputs instead of generated text, supporting MLX, Core AI, llama.cpp, Jev, and Ollama backends.
Unsloth AI@UnslothAIOfficial1 more sourceUnsloth shows how to run Mistral's Le Chonk locally with 1-bit GGUF
Unsloth says its 1-bit GGUF version of Mistral Le Chonk, a 1 trillion parameter model, runs on a Mac Studio and retains 91% top-1 accuracy. The guide lists hardware requirements from 256-270GB for 1-bit to 1.2TB for 8-bit, and the smallest 1-bit quant, UD-IQ1_S, uses about 233GB of disk space.
Industry news
Rohan Paul@rohanpaul_aiXAnthropic accounted for 34.5% of SpaceX's Q3 revenue as its top customer
Rohan Paul relays a Bloomberg report that Anthropic accounted for 34.5% of SpaceX's $14.575B third-quarter revenue, making it SpaceX's largest customer. The payments rent Nvidia GPU capacity at SpaceX's Memphis data centers. Anthropic's S-1 puts the SpaceX arrangements at up to $120.5B through 2029, and says either side can terminate on 90 days' notice after initial periods ending in January and March 2027.
Accenture employee's personal Claude agent accessed a secure Pentagon network
According to sources cited by Wired, an Accenture employee connected a secure Pentagon system to a personal Claude Code account, and the agent ran on the network for two months before being identified. The agent reportedly sent login credentials and a map of the network to a high-risk third-party website, and Accenture says the allegations were investigated and found to be without merit.
The Verge · AINewsAnthropic's S-1 economics: margins, compute commitments and revenue questions
Joey Brookhart of SemiAnalysis discusses Anthropic's S-1 filing with Nilay Patel on Decoder. He says the filing shows inference gross margins of 85 percent, a 1.5 to 2 trillion dollar valuation range, and $70 billion-plus in annual recurring revenue, with coding as the main driver. He also says the company's compute commitments and its training spending will need continued revenue growth to be covered.
Alphabet shares fall nearly 10% premarket after weak search and cash flow results
Alphabet shares fell sharply in premarket trading after third-quarter results showed search advertising revenue of $63.35 billion, up 12%, which missed analyst expectations. The company also reported negative free cash flow of $9.07 billion, far below the $2.70 billion analysts expected, while it kept its full-year capital expenditure guidance at $195 billion to $205 billion. Microsoft rose nearly 4% premarket after reporting $45.2 billion in capital expenditures and $18.63 billion in free cash flow.
Papers & research
Epoch AIOfficial1 more sourceOpen-weight model GLM 5.3 turns unreleased Firefox fixes into working exploits
Epoch AI reports that GLM 5.3, an open-weight model from Z.ai, built working exploits against six Firefox releases from September 17 to October 26 using security fixes published before users received them. Each exploit let a visited webpage read a local file and send it to an attacker, and the marginal cost was about $1,000 per Firefox version. Mozilla shipped Firefox 159.0 on October 27 with a patch for a separate, unpublished vulnerability the model also used. Epoch AI could not produce a comparable exploit against Google Chrome.
Tips & takes
Don't Worry About the Vase (Zvi Mowshowitz)BlogZvi Mowshowitz's weekly AI roundup covers Anthropic's IPO filing and a Slack breach
Zvi Mowshowitz's weekly newsletter reports that Anthropic filed its S-1 for an IPO, with share price and some figures still missing. It also reports that white-hat researchers at Cathedral reached Anthropic's Slack through third-party vulnerabilities, which Anthropic says was out of scope of its safe harbor policy and did not expose model weights. The post also covers Google's Gemini 4 Carbon and OpenAI's firing of three safety researchers, whose accounts the author says are contested.