Mistral releases Large 4, a 1-trillion-parameter open model
France's Mistral released the weights of Large 4, a roughly 1 trillion-parameter model, and says it is the best open model from the West for coding.
French AI company Mistral released the weights of Mistral Large 4 on Thursday, 29 October 2026, reported by Fortune and Trending Topics. It is a model of about one trillion parameters, the internal numbers a model learns, and Mistral calls it the best open model from outside China for coding. Independent scores are less flattering: it ranks eighth among open models. For companies that want to run AI on their own servers, the gap between the claim and the ranking is the story.
What Mistral released
"Open-weight" means anyone can download the trained model from Hugging Face, run it and adapt it, though the training data and code are not necessarily shared. A limited preview went to select partners on 6 October. Fortune says Mistral trained the model on its own 10-megawatt computing cluster in France. Trending Topics says it is a "mixture of experts" design: only about 52 billion of the one trillion parameters work on any one request, like a large firm where only the relevant department handles each query. That keeps running costs lower than the headline size suggests. It handles text and images and supports more than 160 languages.
What Mistral claims
Pierre Stock, Mistral's vice president of science, says Large 4 is the best open-weight model for coding and for cyber defence. Mistral says it narrowly tops other open models on Artificial Analysis' Coding Agent Index, ties China's GLM-5.3 on Terminal-Bench 4.0, a test of coding agents working in a command line, and uses 47% fewer tokens per task than GLM-5.3. Tokens are the small chunks of text a model reads and writes, so fewer tokens means lower cost. Mistral reports 41% on Terminal-Bench, up from 28% at the preview.
What independent testing shows
Trending Topics reports that on Artificial Analysis' overall Intelligence Index Large 4 scores 26, eighth among open models and the highest from outside China. Seven Chinese models rank above it, led by DeepSeek V4.1 Flash and GLM-5.3 at 32 points each. Meta's Muse Spark Open scores 25. On the Coding Agent Index the same report puts Large 4 second, one point behind GLM-5.3, and says Artificial Analysis measured 25% on Terminal-Bench, not 41%. So Mistral's "narrowly tops" claim and the independent ranking differ. We cannot say which is right; they may use different settings. Fortune adds that proprietary models from OpenAI and Anthropic remain ahead overall.
Cost and licence
Trending Topics gives API prices of about 0.63 euro per million input tokens and 1.89 euro per million output tokens, roughly half the preview price. Fortune says that at preview stage Artificial Analysis found the model cost more than four times as much per task as open models of similar intelligence, so the cheaper full-release price matters. Trending Topics says the licence is a modified MIT licence: commercial use is allowed, but a company offering it as a hosted service needs a separate licence from Mistral. Fortune does not name a licence, so check the Hugging Face page before building on it.
What it means for readers
Stock pitched the model to regulated sectors such as finance and healthcare. His argument is that access to closed models "can be cut off", citing a June US export-control directive that forced Anthropic to suspend two of its models for nearly three weeks, per Fortune. He also questioned whether Chinese open models can be trusted. These are Mistral's arguments, not independent findings.
For Indian banks, NBFCs and other companies that must keep data on their own servers, there is now a large non-Chinese open option. Teams should test it on their own tasks, since the vendor's and the independent benchmark figures differ, and check hardware needs, which neither report states.
What to watch next
Watch for independent re-tests that settle the Terminal-Bench gap and for how many cloud providers offer the model. Also watch whether Mistral's licence terms change how companies deploy it as a service.
Key takeaways
- Mistral released the weights of the roughly one-trillion-parameter Large 4 on 29 October 2026, after a preview that began on 6 October.
- Mistral says it leads open models on coding, but Artificial Analysis ranks it eighth overall among open models, behind seven Chinese ones.
- The Terminal-Bench result is 41% by Mistral's count and 25% by Artificial Analysis' measurement, so treat vendor benchmarks with care.
- The licence allows commercial use but requires a separate agreement to sell it as a hosted service.
Sources
- [Fortune: Mistral releases new model Large 4 and claims top open model for agentic coding and cyber defense](https://fortune.com/2026/10/29/mistral-releases-new-model-large-4-and-claims-top-open-model-for-agentic-coding-and-cyber-defense/)
- [Trending Topics: Mistral Large 4 goes open weight, ahead of Meta, behind China](https://www.trendingtopics.eu/mistral-large-4-open-weights-artificial-analysis-meta-china-ranking/)
Tags: Mistral Open-weight models AI models Europe Artificial Analysis