Mistral AI has released the full weights of Mistral Large 4, a one-trillion-parameter Mixture-of-Experts model that anyone can now download, fine-tune and run on their own hardware. The Paris-based lab first showed the model in a public preview earlier in October; the open-weight release, reported by TechCrunch on October 29, comes with an unusually specific pitch: agentic coding and “sovereign” cyber defense.
It is a rare trillion-parameter open-weight release from a European lab, and it lands in a market where many of the strongest open models have come from China. Here is what was released, what the license actually allows, and what it means if you build software.
Key takeaways
- 1 trillion parameters, sparse: a Mixture-of-Experts design with roughly 52B parameters active per token, according to the Hugging Face repository name and developer coverage.
- Open weights, with one catch: a “modified MIT” license allows free commercial use, but companies selling the model as a hosted service need a separate license from Mistral.
- Built for agents and security: Mistral reports leading open-model scores on Terminal-Bench 4.0 (41%) and Cognition’s FrontierCode 2.0 (19.7%), plus a 51% Artificial Analysis Cyber Index score.
- Cheap to call: API pricing is $0.70 per million input tokens and $2.10 per million output tokens.
- All benchmarks are self-reported; a full technical report is still “coming soon.”
What happened
Mistral Large 4, nicknamed “Le Chonk” on Mistral’s own announcement page, is a hybrid model: one set of weights handles both quick instruction-following and longer reasoning. Developers choose between three reasoning levels. Low is for simple tasks, Medium for everyday work, and High, the default, for problems that need careful multi-step thinking. The model accepts multimodal input and, according to Mistral, is fluent in more than 160 languages.
Mistral says the final release had about three more weeks of reinforcement learning than the preview, and that those gains had not plateaued. The company also stresses provenance: the model was “trained from scratch on European infrastructure,” a point aimed squarely at governments and regulated businesses that want to know where their AI came from.
The weights are on Hugging Face as mistralai/Mistral-Large-4-1T-A52B. The model is also available through Mistral’s API and Studio console, OpenRouter, Microsoft Foundry and NVIDIA-hosted endpoints, with day-one support for self-hosting on vLLM and SGLang. Third-party coverage, including AlphaSignal, reports a context window of up to 1 million tokens and an NVFP4 quantized checkpoint; Mistral’s announcement post does not state the context length, so confirm it on the model card before you plan around it.
The numbers Mistral is pointing to
Mistral is not claiming to beat the best closed models across the board. Its argument is narrower: Large 4 is the strongest open-weight option for agentic coding and security work, and it gets there efficiently. The headline claims, all from Mistral’s own testing:
- Terminal-Bench 4.0: 41%, up 13 points from the preview, while producing about 30% fewer output tokens.
- Artificial Analysis Coding Agent Index: 48%, effectively tied with China’s GLM-5.3, but using 47% fewer tokens per task on average.
- Human preference: in a blind coding evaluation by Surge, testers preferred it over GLM-5.3, Kimi K3 and DeepSeek V4.1 Flash. Only Anthropic’s Claude Opus 5.5 ranked higher among the models evaluated.
- Security: 51% on the Artificial Analysis Cyber Index, 97.1% attack resistance on the B3 agent security benchmark, and a 96.6% refusal rate on malicious cyber requests.
Token efficiency matters more than it sounds. For agents that run dozens of steps per task, a model that reaches the same answer with half the tokens is roughly half the cost and noticeably faster, even before you factor in Mistral’s low per-token price.
Why it matters
This year, many of the most capable open-weight models have come from Chinese labs, with models such as GLM, Kimi and DeepSeek setting the pace. Mistral Large 4 gives Western companies and European governments a comparable option whose training lineage they can document, which is increasingly a procurement requirement rather than a nice-to-have.
The cybersecurity framing is the other big story. Mistral’s VP of Science Pierre Stock told TechCrunch the appeal is “access that can’t be cut off,” because an organization running the weights on its own servers does not depend on a vendor’s API staying available. His argument is that defenders need strong AI tooling quickly, since attackers are already using it. That case has real weight: this same week, CrowdStrike described a state-linked campaign that used an AI agent framework to target more than 200 organizations.
It also continues a trend we covered with Google’s Gemma 4: big labs now treat open weights as a strategic product, not a side project.
What it means for developers and businesses
Start with the API. Running a trillion-parameter model yourself needs serious multi-GPU server hardware, even though only a fraction of the parameters fire per token, because all of the experts still have to sit in memory. For prototyping, the hosted API or OpenRouter is the practical route, and at $0.70/$2.10 per million tokens (with cached input at $0.07) it is inexpensive by frontier-model standards.
Self-host when control matters. If you work with sensitive code, regulated data or air-gapped environments, the open weights plus vLLM or SGLang support make on-premises deployment realistic. This is not a laptop model; if you want something you can run on your own machine, our guide to local LLMs in 2026 covers smaller options.
Read the license before you build a product on it. The “Mistral Ecosystem Open-Weight” license permits use, modification, fine-tuning and commercial use. But if your business is offering the model itself as a hosted service, as a Model-as-a-Service provider, you need a separate commercial agreement with Mistral first. Using it inside your own app or internal tools is fine; reselling inference is not, without that deal.
Test it on your own agent tasks. Mistral’s numbers are promising but self-reported. Run it against your existing coding-agent or security workflows and compare cost per completed task, not just accuracy.
What to watch next
- The technical report. Mistral says it is coming soon. It should confirm the exact architecture and context length and show how the benchmarks were run.
- Independent evaluations. Watch for third-party leaderboard results and red-team testing of its cyber capabilities.
- Fine-tunes and quantizations. Open weights usually lead to community variants within weeks; smaller distilled versions would widen who can run it.
- The response from Chinese labs. GLM, Kimi and DeepSeek have shipped new open models quickly all year, and the open-weight lead could change hands again soon.
The short version: Mistral Large 4 is a credible, cheaply priced open-weight model for coding agents and security teams, with a license that is generous for builders and restrictive only for those who want to resell it. The independent benchmarks will show whether it is as strong as Mistral says.