Mistral AI released the weights of Mistral Large 4 on October 29, 2026, making its largest model available for download and self-hosting. The French company calls the model Mistral Large 4, and it carries the internal nickname Le Chonk. The release followed a public preview that began on October 6, 2026, during which Mistral said it continued reinforcement learning and revised the final checkpoint.
The model is natively multimodal, which means it accepts text and image input and produces text output. It has about 1 trillion total parameters. Only a small share of those parameters runs for each token, The model card lists 49 billion active parameters per token, excluding embeddings and output layers, and 52 billion when those layers are included. It uses 128 routed experts, with four selected for each token, plus one shared expert.
Mistral published the weights on Hugging Face under a modified MIT license, described as the Mistral Ecosystem Open-Weight License. The terms allow use, modification, fine tuning and commercial use, with one condition. A business that offers the model as a hosted service must obtain a separate commercial license from Mistral before doing so commercially.
Mistral Large 4 coding results
Mistral reported that Mistral Large 4 leads Cognition’s private FrontierCode 2.0 benchmark, which tests end to end code quality and whether a repository maintainer would accept an agent’s change. The reported score was 19.7 percent, ahead of listed open weight models including GLM-5.3, Kimi K3 and DeepSeek V4.1 Flash. The company also reported 48 percent on the Artificial Analysis Coding Agent Index, described as effectively tied with GLM-5.3, while using 47 percent fewer tokens per task on average.
On Terminal-Bench 4.0, Mistral reported a score of 41 percent, up 13 percentage points from the public preview. The company said that result leads open weight models on that test and that output token use fell by 30 percent. In Surge Code blind pairwise human preference tests cited by Mistral, the model was preferred over GLM-5.3, Kimi K3 and DeepSeek V4.1 Flash, and ranked behind only Claude Opus 5.5 among the models tested. These are company reported results.
Mistral Large 4 supports a context window of 1 million tokens, subject to serving memory and runtime limits. Mistral said it trained the model from scratch on 3,800 Nvidia Grace Blackwell graphics processing units in Europe, and that it is fluent in more than 160 languages, including every official language of the European Union. It offers three reasoning modes, low, medium and high, with high as the default.
Mistral Large 4 security work and limits
Mistral is pitching Mistral Large 4 at cyber defenders. The company reported a 51 percent score on the Artificial Analysis Cyber Index, which measures discovery of vulnerabilities in a codebase, reproduction and validation of those flaws, and patching without breaking existing functions. It also reported 93 percent on Cybench, a set of 40 challenges drawn from security competitions, . Mistral said the model handles legitimate defensive security tasks without unnecessary refusals.
The company reported an average refusal rate of 96.6 percent on malicious cyber requests across the cyber related prompts of JailbreakBench, StrongREJECT and AgentHarm. It described that rate as the highest it measures among open models. Mistral also said it observed a potential for uplift to conduct real attacks during the preview period, and it is changing its guardrails in response. A specialized variant built for penetration testing and red teaming will go only to vetted enterprise design partners, including Orange Cyberdefense, under contract.
Open weights cannot be recalled once they are downloaded, and a determined user can try to fine tune away refusal behavior. Mistral said its guardrails are built to distinguish defensive work from malicious offensive requests. The company tested the model with CrowdStrike, Cloudflare, Rubrik, Zscaler, Fortinet and Orange Cyberdefense on enterprise security workloads.
Mistral Large 4 access and pricing
Through Mistral’s own interface, pricing is $0.70 per million input tokens, $0.07 per million cached input tokens and $2.10 per million output tokens, in the company’s model documentation. Microsoft is the first cloud provider to carry Mistral Large 4 through Foundry, and Nvidia offers it as a NIM container. The model can run on as few as four Nvidia B200 or B300 graphics processing units, a modest footprint for a model of this size.
Mistral said a full technical report is coming soon. The company also said Mistral Large 4 will serve as the foundation for a new generation of specialized and optimized Mistral models, pointing to recent releases such as Voxtral and Timestral. For now, the release gives teams that want to run a large model on their own infrastructure a new option.