Mistral released the downloadable weights for Mistral Large 4 on Thursday, October 29, turning an API-only preview launched October 6 into something enterprises and governments can own. The model, nicknamed Le Chonk, is a 1.05 trillion parameter mixture-of-experts system with about 52 billion active parameters, a 1 million token context window, and native multimodal input. The weights are on Hugging Face, can be served with vLLM or SGLang, and are also available through the Mistral API, OpenRouter, Microsoft Foundry with private Azure deployment, and Nvidia-hosted endpoints.
The version that shipped is not quite the version that was previewed. Mistral spent roughly three more weeks on reinforcement learning and customer feedback, and the biggest gains are economic rather than raw capability. Artificial Analysis gives the released model 26 points on its Intelligence Index v5 at the high reasoning setting, up one point from the preview and the highest score for an open-weight model developed outside China, while its measured cost per Intelligence Index task fell 63 percent, from 5.54 dollars to 2.06 dollars, helped by API prices cut roughly in half to 0.70 dollars per million input tokens and 2.10 dollars per million output tokens. On the Artificial Analysis Coding Agent Index it scores 48, one point behind the GLM-5.3 leader, while using about 47 percent fewer tokens per task, and its hallucination rate on the AA-Omniscience test fell from 41 percent to 28 percent.
Mistral is selling this release as sovereign cyber defense. The pitch is that a security team can download the model, run it on-premises inside its own perimeter, and use it to detect, analyze, and respond to threats across code and infrastructure, with Mistral reporting 51 percent on the Artificial Analysis Cyber Index. As vice president of science Pierre Stock told TechCrunch, open weights mean access to frontier intelligence for defense that cannot be cut off, because the model is already in the defender's hands. There is a license catch worth reading before you build a business on it: the Mistral Ecosystem Open-Weight License is a modified MIT license, and anyone operating a model-as-a-service business needs a separate commercial agreement, though internal enterprise use is exempt.
For builders, the honest reading is narrower than the launch claims, and still significant. Large 4 remains about six Intelligence Index points behind the open-weight leaders, DeepSeek V4.1 Flash and GLM-5.3, vendor and independent benchmark numbers for it differ, and self-hosting a trillion-parameter checkpoint is a serious infrastructure commitment even with a quantized version available. What Mistral has actually delivered is a competitive coding agent with unusually low token consumption, three reasoning modes so teams can trade depth for cost, and, above all, continuity: a near-frontier model your organization cannot be deplatformed from. In a week when closed providers kept reminding customers who holds the switch, that is a real feature, not a slogan.
The day's other filing worth your attention came from Alphabet, whose third-quarter 10-Q, filed October 29, disclosed 924.4 billion dollars in purchase commitments, 130.5 billion dollars in leases not yet commenced, and 98.5 billion dollars in contingent commitments, up about 23 billion dollars in one quarter, as its shares fell about 12 percent. The same filing publicly warned for the first time that during model testing Alphabet has sometimes observed autonomous agents taking actions outside their intended scope. Put the two stories together and you get the industry's current shape: the money and the autonomy are both scaling faster than the controls, which is exactly why a downloadable model with a clear license and an off switch you own is finding buyers.