AI · Published
Mistral AI has released the weights for Mistral Large 4, a trillion-parameter model that the French company says is the best open model for agentic coding and cyber defence. TechCrunch and Fortune both reported the release on 29 October 2026. The weights are available on Hugging Face and through Mistral's API. Mistral first introduced the model in a preview on 6 October, according to Fortune.
Most of Mistral's claims focus on efficiency. TechCrunch reports that the company says Large 4 reaches state-of-the-art performance using a fraction of the compute and capital that closed labs commit. Mistral says the model is comparable to GLM-5.3 and uses 47 percent fewer tokens per task. It also says Large 4 outranks every other open model on Artificial Analysis's Coding Agent Index. TechCrunch does not publish benchmark numbers for that ranking.
Fortune, citing Mistral's benchmark figures, reports that Large 4 ties GLM-5.3 on Terminal-Bench 4.0. Its score rose from 28 percent to 41 percent after three more weeks of reinforcement learning following the October preview. Fortune also reports that Artificial Analysis rated the preview version roughly on par with GPT-6 Luna and DeepSeek V4.1 Flash, but at more than four times the cost per task of similar open models. The benchmark results are the company's own figures, and the Artificial Analysis rating applies to the preview rather than the released weights.
Fortune says Mistral's vice president of science, Pierre Stock, described training on a 10-megawatt cluster in France. TechCrunch quotes Stock on cyber defence and on-premises deployment, which points to security teams that want to run a large model on their own hardware instead of sending data to a hosted service. Neither report says how many organisations have started using the model.
The company is well funded for this push. Fortune reports that Mistral raised a €3 billion Series D in September, led by Samsung, at a valuation above €21 billion. The reports do not say how much of the training cost that round covered, so the funding alone does not show how the model was paid for.
Several questions remain open. Neither report names the licence that governs the weights, so the terms for commercial use are not confirmed by the coverage reviewed for this article. The benchmark figures have not been independently reproduced in the sources read. It is also unclear how the 41 percent Terminal-Bench score compares with results other labs obtain under the same test conditions. Readers should check Mistral's own model card and independent evaluations as they appear.
The release matters because it puts a model of this size in the hands of developers who can inspect and run it without a paid API. Open-weight releases from Chinese developers have set much of the recent pace in this category, so a Western open model that claims parity on coding and agent tasks is a useful test of how quickly that gap is closing. For now, the claims are Mistral's, and the independent results will show whether they hold.
Sources
This article was researched and written by an AI system from the sources listed above, under SavvyToolz's sourcing and freshness rules. Read the editorial policy.