← Dailies

Daily · 2026-10-30 00:00 UTC

Mistral opens trillion-param weights; DOJ eyes Nvidia-Hugging Face

Mistral releases the weights for Mistral Large 4, the Justice Department probes Nvidia's $12.93B Hugging Face purchase, and Cohere opens a Series E targeting C$7B.

Source ↗

The open-weight camp made its biggest argument yet, Washington started asking what Nvidia would own if it buys Hugging Face, and a Canadian lab went looking for countries that do not want to pick between American and Chinese AI. Here are the ten signals that mattered in the last 24 hours.

1. Mistral releases the weights for Mistral Large 4, a trillion-parameter model — Nicknamed "Le Chonk," the model was first introduced earlier this month; the weight release means businesses and governments can download and deploy it themselves. Mistral's case is control: organisations that want to keep data and systems in-house, with cybersecurity as the sharpest example. CEO Arthur Mensch framed AI as something that should "be a tool for empowerment, not dominance," and the company says it reaches state-of-the-art performance at "a fraction of the compute and capital committed by closed labs." Mistral says ML4 is comparable in performance with Chinese open-weight competitor GLM-5.3. TechCrunch · Oct 29
2. The Justice Department is scrutinising Nvidia's $12.93B Hugging Face deal — Officials cited by Reuters are asking whether the purchase would give Nvidia control of a key hub for distributing AI models and hurt competitors. Nvidia expects the deal to close in the first half of 2027 and says Hugging Face will keep supporting partnerships across chip, cloud and AI providers. The wider antitrust probe covers Nvidia's acquisitions, investments, sales practices and possible pressure on customers, including its SchedMD and Run:ai deals and a licensing arrangement with Groq, and whether customers were offered more chips for also buying networking gear and Groq processors. France's competition authority is separately investigating dependence on CUDA and Nvidia's stakes in cloud providers such as CoreWeave. The Decoder · Oct 29
3. OpenAI doubles usage limits across its paid tiers — Work and Codex quotas on the Pro 100 and Pro 500 plans both get twice the previous allotment. On Pro 200, OpenAI doubled the lower quota and reversed a planned cut after user pushback; Plus is unchanged. New subscriptions include the new quotas and existing customers get them automatically. OpenAI cautions that a doubled quota does not mean twice as many completed tasks, since consumption varies by model, task type and settings, and says it will keep shipping efficiency gains — pointing to its recently released 6.1 models Terra and Sol. The timing lands just ahead of Anthropic's IPO as both companies fight for enterprise business. The Decoder · Oct 29
4. Cohere opens a Series E targeting up to C$7B (about US$5B) — The Canadian company announced the initial close in a blog post that rejects what it calls "frontier fatalism" — the belief that AI will be monopolised by a handful of US and Chinese giants. Cohere is positioning itself as a sovereign alternative for other countries, and says it will use the funds to build "something increasingly essential in AI, a frontier company" rather than depend on the incumbents. PYMNTS · Oct 29
5. Meta quietly launches an AI fashion app called Style Sketch — The app connects to Instagram to build a virtual closet from photos you have already posted, or lets you add items by snapping a photo, recording a short video or uploading from your camera roll. It then suggests outfit combinations and lets users try them on a "realistic model matched to your body shape and skin tone." It is available on iOS and Android in Indonesia, Singapore and Thailand; Meta did not respond to a request for comment and has not said whether it will launch more widely. TechCrunch · Oct 29
6. Gemini can now edit Docs, PDFs, spreadsheets and slides inside the chat — Google added preview and editing for most office file formats directly in the Gemini chat window on the web, so editing a file Gemini generated no longer means leaving the app. The Verge · Oct 29
7. GM to pay $220M to settle claims it sold drivers' data — A coalition of 36 state attorneys general secured the settlement, which resolves claims that GM illegally sold driver and location information to data brokers. The agreement also requires GM to limit the data it collects. The Verge · Oct 29
8. What it is like to work when your co-workers are AI agents — The Guardian reports from San Francisco startups where agents fill roles from executive assistant to chief of staff: one founder at the startup Anything talks daily with an assistant called Gwen, a coding partner called Cody and a search agent called Betsy. Companies are boasting about doing more with less, and the piece documents the tension that creates with the humans working beside them. The Guardian · Oct 29
9. Prime Intellect makes DeepSeek's Mega MoE kernel trainable — Prime Mega MoE extends DeepSeek's fused Mixture of Experts kernel with a fast, overlapping backward pass so it can be used for training rather than inference alone: up to 3.11x faster MoE kernels and up to 1.34x faster end-to-end training in Prime-RL. Prime Intellect · Oct 29
10. Baseten releases joey-1, a 2B open-weights decision model — Post-trained from Qwen3.5-2B, joey-1 classifies text, routes requests and scores content against a rubric, returning typed answers an application can use directly: choices, yes/no probabilities and scores. Several questions can ride in one request, and questions about the same state reuse its input prefix so they cost less together than apart. It is available through Baseten's Model APIs with weights open on Hugging Face. Baseten · Oct 29

🧪 Try This Next

• 🔧 Put joey-1 in front of a larger model as a router or classifier — typed outputs beat parsing free text from a frontier model.
• 🎥 Read the Guardian's agent co-worker piece before pitching "AI headcount" to a team that has to work alongside it.
• 📝 Park the Mistral Large 4 weights as your self-hosted fallback and check the licence against your compliance rules first.

Content on Anagnorisis is summarized, paraphrased, and editorialized from publicly available sources for length and clarity. Original sources are linked where available. All trademarks belong to their respective owners.