Mistral has now done what it said it would: the Paris-based open-weight lab published the downloadable files for Mistral Large 4, its 1 trillion-parameter model, and released a retrained version that scores higher than the preview did. Until this announcement, the model could only be reached through Mistral’s own servers. Anyone can now pull the weights from Hugging Face and run them on hardware they control, using the open-source serving tools vLLM or SGLang.
The license answers a question the preview left open. The terms are a modified MIT license that Mistral gives its own name, and they let you run the model, change it, retrain it, and sell products built on it at no charge. The single restriction hits companies that run a Model-as-a-Service business: if you resell access to the model by API, you need a separate commercial deal with Mistral. For everyone else, including a bank or a software vendor running it internally, the weights come without a fee.
On coding, Mistral credits three more weeks of reinforcement learning, the training stage where a model is rewarded for solving tasks. According to Mistral AI’s announcement post, the model scores 19.7 percent on FrontierCode 2.0, a private test from Cognition that asks whether a repository maintainer would accept the agent’s change. The next score on Mistral’s own chart is 19.3 percent, so the lead is a fraction of a point. Artificial Analysis, an outside benchmarking firm, runs a coding-agent index where the model posts 48 percent, effectively level with GLM-5.3, the strongest open rival, while using 47 percent fewer tokens per task. On Terminal-Bench 4.0 it gained 13 points over the preview to reach 41 percent. In blind comparisons run by Surge Code, graders picked its answers over those from GLM-5.3, from Kimi K3, and from DeepSeek’s V4.1 Flash. Only Claude Opus 5.5 ranked above it.
The token figure is the one a budget owner will care about, because a coding agent bills by the amount of text it reads and writes. Mistral prices the model at $0.70 per million input tokens and $2.10 per million output tokens, with cached input at $0.07. It also added three reasoning settings, low, medium, and high, with high as the default, so a team can pay for less thinking on easy jobs.
On security, Mistral is franker than most vendors. Large 4 scores 51 percent on Artificial Analysis’s cyber index, a test of finding, reproducing, and patching vulnerabilities, and the same chart lists two other models above it at 56 and 57 percent. Mistral says the model refuses 96.6 percent of malicious cyber requests on average, the highest it measured among open models. Yet the company also writes that live traffic showed the model declining black-box attacks on real targets while still carrying potential for uplift to a real attacker, and that it is tightening its guardrails as a result.
Open weights cannot be recalled, so guardrails that Mistral adjusts on its own API do not follow a copy sitting on someone’s server. A separate variant built for penetration testing and red-teaming goes only to vetted enterprise partners such as Orange Cyberdefense, under contract.
Distribution is already wide. Microsoft says Foundry is the first cloud to offer the model, NVIDIA hosts it as a ready-to-serve container, and OpenRouter and Mistral Studio carry it too. The announcement is a company post, and the benchmark numbers in it are Mistral’s reporting of third-party tests, with no independent reruns attached.
The weights close that gap for anyone willing to do the work. A team can test the 47 percent token claim on its own repositories, and Mistral says the technical report explaining how it got there is still to come.
Reported by Mistral AI in its company blog post “Le Chonk: Built for Coders, Trusted by Defenders”, which carries no publication date.
