Mistral released the weights for Mistral Large 4 on October 29, 2026, turning its trillion-parameter open model into a checkpoint that enterprises and governments can download, run and own. The French company first previewed the model, nicknamed Le Chonk, on October 6, and said it has since applied preview feedback and continued scaling reinforcement learning. The release combines a large mixture-of-experts architecture, native multimodal capabilities and a security focus that extends well beyond generating text.
Mistral Large 4 contains roughly 1.05 trillion total parameters, while activating about 52 billion parameters for each token. That design is intended to provide broad capability without the compute demands of activating every parameter for every request. The weights are published on Hugging Face under the Mistral Ecosystem Open-Weight License, a modified version of the MIT license. Users may use, modify, fine-tune and commercialize the model, although a business operating a model-as-a-service offering must first obtain a separate commercial license from Mistral.
Access is also available through the Mistral API and Mistral Studio, OpenRouter, Mistral’s Vibe Code coding assistant, NVIDIA-hosted endpoints using an optimised NIM container, and Microsoft Foundry. Microsoft corporate vice president of Foundry Tina Schuchman said the company is the first cloud provider to bring Mistral Large 4 to customers, with private deployment available through Foundry for Azure developers. API pricing is 0.70 dollars per million input tokens, 0.07 dollars per million cached input tokens and 2.10 dollars per million output tokens.
Mistral Large 4 targets realistic coding work
Mistral reported that Mistral Large 4 leads Cognition’s private FrontierCode 2.0 evaluation with a score of 19.7 percent, ahead of listed open-weight models including GLM-5.3, Kimi K3 and DeepSeek V4.1 Flash. The benchmark asks whether a repository maintainer would merge an agent’s proposed change, which is a more practical test than checking whether an automated test suite passes. The model also scored 48 percent on the Artificial Analysis Coding Agent Index, a result Mistral described as effectively tied with the leading open-weight model, GLM-5.3, while using 47 percent fewer tokens per task on average.
On Terminal-Bench 4.0, Mistral said the model improved by 13 percentage points over its public preview to reach 41 percent. That put it ahead of the other open-weight models in the company’s comparison while reducing output tokens by 30 percent. In blind pairwise human-preference evaluations called Surge Code, Mistral said Mistral Large 4 was preferred over GLM-5.3, Kimi K3 and DeepSeek V4.1 Flash. It ranked behind only Claude Opus 5.5 among the models evaluated.
These figures are company-reported and should not be read as proof that the model will perform equally well in every repository or production environment. Coding agents still depend on tool access, context quality, testing discipline and the clarity of the task. Even so, the combination of high scores and lower token consumption points to a practical advantage: teams may be able to run more attempts within a time or cost budget without adopting a fully closed system.
Mistral also introduced three reasoning modes. Low is intended for simple tasks, medium supports everyday work, and high is the default for difficult requests that require deeper reasoning. Developers can trade cost for depth when an agent encounters an ambiguous codebase, a difficult bug or a multi-stage planning problem.
Cyber defense is the central enterprise pitch
The cybersecurity results are the strongest part of the launch for security teams. Mistral reported a 51 percent score on the Artificial Analysis Cyber Index, which tests whether a model can find a vulnerability in a codebase, reproduce and validate it, then patch it without breaking existing functionality. In Mistral’s results, Mistral Large 4 handled legitimate defensive work while refusing malicious requests. The company reported an average refusal rate of 96.6 percent on malicious cyber prompts across JailbreakBench, StrongREJECT and AgentHarm, calling it the highest rate it measures among open models.
That distinction between defensive and offensive use is important because a model that refuses every security request is not especially useful to defenders. The benchmark result suggests Mistral is trying to make the model useful for tasks such as reviewing source code, reproducing a flaw and producing a remediation, while blocking requests to build black-box attacks against live targets. Mistral disclosed that it monitored cyber-related activity in live traffic after the preview. The company said it saw the model decline requests for black-box attacks against live systems, but it also concluded that the model could potentially help users pursue real attacks. Mistral is modifying its guardrails and has promised further updates.
Enterprise design partners, including Orange Cyberdefense, are receiving preview access to a specialised version tuned for penetration testing, vulnerability assessment and red teaming. Contractual limits require testing against owned or explicitly authorised systems in controlled environments. The release also included evaluation statements from Booz Allen Hamilton, Cisco, Cloudflare, CrowdStrike, Elastic, ENISA, Fortinet, NVIDIA, Orange Cyberdefense, Rubrik and Zscaler.
Brad Medairy, president of national cyber at Booz Allen Hamilton, said tests showed the model matching the vulnerability-detection recall rates of other frontier cyber models against the firm’s internal benchmark for source-code security analysis. That claim is meaningful because an internal benchmark can reflect workflows that public tests miss, but it remains an early signal. Teams still need to test the model against their own code, validate proposed patches and establish clear human approval gates.
Open weights change the deployment equation
The strategic significance is control. Organisations that process sensitive source code or operate important networks often cannot rely on an outside laboratory to retain access to frontier capability. An API provider can change prices, remove features, update a model or restrict use according to its policies. Downloadable weights allow an organisation to host Mistral Large 4 within its own infrastructure and determine how data is handled, what logs are retained and which teams can access the system.
That does not make deployment free. A model with roughly 1.05 trillion total parameters still requires substantial memory, serving infrastructure, monitoring and technical expertise. Organisations with a small security team may find the hosted API more practical, while governments and large enterprises may have the scale and mandate to operate their own environment. Open weights also move responsibility for updates and safeguards to the deployer, rather than leaving every change in the hands of the original lab.
Mistral chief executive Arthur Mensch framed the release as a bet that artificial intelligence should expand human capability rather than reinforce dominance. Pierre Stock, Mistral’s vice president of science, said open weights provide access that cannot be cut off by definition and allow deployment on premises. For incident response, he said the model could help identify, coordinate and respond to major security incidents in real time.
The same day, Mistral opened its Responses API to public preview. The endpoint is designed as a drop-in replacement for OpenAI’s agent format and returns typed items separately rather than combining prose, reasoning and tool calls into one message. This gives orchestration software cleaner ways to inspect tool calls and process outputs. The API is stateless, so applications must resend the context they need, but that design can make agent pipelines more portable and predictable.
OpenAI also reorganised its Model Spec around faithfulness to human intent, harm minimisation, honesty and transparency, human agency and care. Its new chain of command states that models should not pursue revenue, self-preservation or additional compute as ends in themselves. The parallel announcements underline a broader change in AI infrastructure. Closed services still set the pace on many benchmarks, while open-weight systems are becoming credible alternatives for organisations that value deployment control and technical ownership.
What Mistral Large 4 changes for defenders
Mistral Large 4 is not automatically safer because its weights can be downloaded, nor is a benchmark lead a security guarantee. The release does, however, give defenders a large multimodal model they can inspect, adapt and operate in environments where external APIs may be unacceptable. Its strongest case is not that it replaces every security analyst or closed frontier model. It is that teams can test a capable agent against proprietary code, preserve control over sensitive material and decide for themselves where human review belongs.
For organisations considering adoption, the sensible first step is a controlled evaluation. Security teams should test the model on representative repositories, compare its findings with established scanners and manual review, measure the cost of each accepted patch and document how the system handles sensitive data. They should also treat the remaining offensive-use risk as an active engineering concern, adding network controls, authorisation checks and human approval before any action reaches production systems. Whether open weights at this scale actually change who can run frontier security models, and whether the benchmark leads hold up outside Mistral’s own testing, will decide how much Mistral Large 4 shifts the open-weight balance.
Mistral Large 4 puts a trillion-parameter model, advanced coding ability and a refusal layer tuned for defensive work under the control of the organisations expected to use it. Whether it becomes indispensable will depend on more than a leaderboard result. The decisive test is whether defenders can run it securely, verify its work and respond faster without losing accountability. Mistral Large 4 gives that effort a new foundation, but the operating discipline around the model will determine whether open weights become a genuine security advantage.