Mistral releases the weights of Large 4: what the numbers do and do not show
What happened
On 29 October 2026 the French company Mistral AI published the weights of Mistral Large 4, its largest model, on Hugging Face. The release date is given by Artificial Analysis’s model listing and by XenoSpectrum’s report; Mistral’s own post describes the release as happening “today” without a date. The model had been available earlier in October as an API preview, and Mistral had said then that weights would follow by the end of the month.
Large 4 is a mixture-of-expertsmixture of experts A model design split into many sub-networks, where only a few of them are switched on for each piece of input, so most of the model sits idle at any moment. model with about one trillion parametersparameter One of the adjustable numbers inside an AI model that training tunes; model size is usually counted in them. in total (Mistral says 1T; DeepLearning.AI’s The Batch reported 1.05 trillion) and about 52 billion active for each tokentoken A small chunk of text, often a word or part of a word, that a language model reads or writes as one unit., according to The Batch and Artificial Analysis. It accepts text and images, has three reasoning settings, and Artificial Analysis lists a context windowcontext window The maximum amount of text a language model can take in and consider at once. of one million tokens. It is released under what Mistral calls the Mistral Ecosystem Open-Weightopen-weight Describes an AI model whose trained parameters are published so anyone can download and run it, even if the training data and code are not. licence, a modified MIT licence that allows commercial use and fine-tuningfine-tuning Further training of an already trained AI model on a smaller, chosen set of examples to change how it behaves. but requires a separate agreement for anyone selling the model as a hosted service. API pricing is $0.70 per million input tokens and $2.10 per million output tokens.
What is claimed vs. what is verified
Mistral’s claims. The company describes Large 4 as the best open-weight model for coding. It reports 41 percent on Terminal-Bench 4.0 (up from 28 percent for the preview), 48 percent on Artificial Analysis’s Coding AgentAI agent An AI system that carries out multi-step tasks on its own, such as running code or using tools, rather than only answering one message at a time. Index, 19.7 percent on Cognition’s private FrontierCode 2.0 benchmarkbenchmark A fixed set of tasks used to score and compare AI systems., and 84.5 percent on BrowseComp. It says the model refuses 96.6 percent of malicious cyber requests and resists 97.1 percent of attacks on the B3 agent security benchmark. These figures appear in Mistral’s own post; several cite outside benchmarks but were reported by Mistral, and none of the refusal or security numbers has been reproduced by an independent group so far.
Independent evidence. Artificial Analysis, a firm that runs its own evaluations, gives Large 4 at its high reasoning setting a score of 26 on its Intelligence Index v5, above the median of 23 for comparable models. XenoSpectrum, citing the same firm, notes that GLM-5.3 and DeepSeek V4.1 Flash each score 32. Large 4 costs less per task than GLM-5.3 ($2.06 against $5.01) but more than DeepSeek V4.1 Flash ($0.60). Artificial Analysis also measures output speed of about 161 tokens per second, well above the class median. XenoSpectrum points out that it lists a FrontierCode figure different from Mistral’s, and that reasoning settings and index versions differ across comparisons, so not every figure is like for like.
An earlier analysis of the preview by The Batch, on 16 October, reported a score of 38 on a previous version of the index and called the model’s cost per task high compared with rivals at the same score. Because the index has since changed version and the model was still in training, those numbers cannot be compared directly with today’s.
Why it matters
Large 4 is among the largest models whose weights anyone can download, and it comes from a European company at a time when the strongest open-weight models have mostly come from Chinese labs. Organisations that cannot send data to an outside API can now run a model of this size on their own hardware, subject to the licence.
The release also tests how labs handle cyber capability in open models. Mistral writes that after the preview it still saw potential for real-world uplift in attacks and is adjusting its guardrails. A less restricted cybersecurity variant remains limited to vetted partners. Safeguards built into a model can be weakened by anyone who fine-tunes the open weights, which is a general limitation of this approach rather than a finding about Large 4.
Open questions and other views
Supporters of open weights argue that releases like this let researchers audit models, let smaller firms and governments avoid dependence on a few providers, and spread the benefits of defensive security tools. Critics of releasing highly capable models argue that refusal training can be removed after release and that cyber uplift, which Mistral itself acknowledges, is hard to take back once weights are public.
On performance, the open question is how Large 4 does on evaluations Mistral did not choose. Independent results so far come mainly from one firm, and the gap with Chinese rivals rests on a single aggregate index.
Sources
- Mistral AI, “Mistral Large 4 (Le Chonk) is now open weight,” company announcement, October 2026. mistral.ai
- Artificial Analysis, Mistral Large 4 model page, accessed 30 October 2026. artificialanalysis.ai
- Y. Kobayashi, “Mistral Large 4 Open Weights Released: Better Coding Efficiency, but Gaps With Chinese Rivals Remain,” XenoSpectrum, 29 October 2026. xenospectrum.com
- DeepLearning.AI, “Mistral’s skills match its global appeal,” The Batch, 16 October 2026. deeplearning.ai