DeAI Daily Brief — 30 October 2026

Today in DeAI: Mistral published Mistral Large 4 weights with a commercial-hosting license catch, the US Justice Department is scrutinizing Nvidia's planned Hugging Face acquisition, and OpenAI doubled Pro usage limits while adding subscription-funded API credits.

Mistral Large 4 weights are out, with a license catch

Mistral AI released the weights of Mistral Large 4 on October 29, a 1T-parameter multimodal model with 49-52 billion active parameters and a 1M-token context window, downloadable from Hugging Face under a modified MIT license. The condition: businesses running the model as a hosted model-as-a-service offering need a separate commercial license from Mistral. Mistral claims the open-weight lead on Cognition's FrontierCode 2.0 at 19.7 percent, while independent Artificial Analysis scoring puts it at 26 points, behind DeepSeek V4.1 Flash and GLM-5.3 at 32 each. Why it matters: the license walls off the hosted-serving market for a frontier-scale model, and every builder should read the actual license file before planning a deployment. (Mistral announcement) — our full coverage

Justice Department scrutinizes Nvidia's planned Hugging Face acquisition

The US Justice Department is examining Nvidia's planned acquisition of Hugging Face along with Nvidia's sales practices, per The Decoder. No complaint or remedy has been filed; this is reported scrutiny, not an enforcement action. Hugging Face hosts effectively every open-weights release covered on this beat, so the practical questions for builders concern packaging, licensing presentation, and discoverability of non-NVIDIA models if the deal ever closes. Why it matters: the hub's neutrality as one-URL distribution for open weights is infrastructure the whole ecosystem quietly depends on. (The Decoder)

OpenAI doubles Pro limits and adds expiring API credits

OpenAI doubled usage limits across its paid plans, from Pro 100 through Pro 500, reversed a planned Pro 200 quota cut after user pushback, and launched expiring API credits for $100, $200, and $500 subscribers, per its help center. The Decoder ties the moves to competition with Anthropic, which added a similar benefit on October 7. Why it matters: subscription-to-API credit transfer is becoming a standard entitlement; if you run agents across both, the credit expiry terms, not the headline amounts, decide what a subscription is actually worth. (OpenAI Help Center)

Baseten ships joey-1, a 2B decision model, at a claimed $0.01 per million tokens

Baseten released joey-1, a 2-billion-parameter open-weight model that returns calibrated probabilities instead of text, with claimed 20 ms P50 latency and $0.01 per million input tokens. Those figures are Baseten's self-report via MarkTechpost until they are checked against the model card and pricing page. The structure is the notable part: an inference provider shipping its own small open model for routing and classification workloads. Why it matters: cheap calibrated scoring is the unglamorous layer most production AI systems lean on, and per-token pricing for a 2B model resets what classification work should cost. (MarkTechPost)

Cohere details Model Vault confidential inference

Cohere published an engineering deep-dive on Model Vault, which runs inference inside confidential CPU and GPU TEE enclaves with end-to-end encryption and remote attestation, so customers can verify that prompts and weights are shielded from Cohere's own operators. The design description is provider-authored; what the attestation actually measures and the enclave's performance cost still need an independent read before the privacy property counts as proven. Why it matters: attestation-based confidentiality is moving from bespoke deployments to productized offerings, and the verifiable parts of these designs are what separate them from a privacy policy. (Cohere blog)

Watching tomorrow

Moonshot AI's Kimi K3.1 Pro weights are due by November 4 and would re-rank the open-weight coding benchmarks Mistral just claimed; the weekend also gives independent evals of Le Chonk time to land.