Mistral has released the weights for Mistral Large 4, the 1-trillion-parameter model it nicknames Le Chonk, which puts a frontier-scale model in the hands of anyone with the hardware to run it. The model went into public preview earlier in October, and on October 29 Mistral published the weights along with updated benchmarks and a new license. If you've been following the open versus closed debate, this release is worth a close look, so below we cover what the model is, what the numbers say, what the license allows and who can realistically run it.
What Mistral Large 4 is
Mistral describes Large 4 as an open-weight hybrid instruct-and-reasoning model with multimodal input. On Mistral's Hugging Face page, it's listed at 1 trillion total parameters with 52 billion active, built on a mixture-of-experts architecture, natively multimodal for text and vision, with a 1 million token context window and support for more than 160 languages. The main repository is Mistral-Large-4-1T-A52B, and there's also a quantized NVFP4 version.
The mixture-of-experts design is why a model this large can still be practical. Only a slice of the network, about 52 billion parameters, is active for each token, so inference costs look more like a mid-sized model than a trillion-parameter one. Mistral also says the model was trained from scratch on European infrastructure, which matters to companies that care about where a model came from.
If the terms here are new to you, our explainer on open-weight vs closed AI models walks through what you get when a lab releases weights and what you still don't get, like training data.
What the benchmarks say
Mistral is pitching Large 4 mainly as a coding and agent model. The company says it leads Cognition's private FrontierCode 2.0 benchmark with a score of 19.7% and reached 41% on Terminal-Bench 4.0, a 13 point jump from the public preview. On the Artificial Analysis Coding Agent Index it scores 48%, which Mistral describes as effectively tied with GLM-5.3, the leading open-weight model there, while using 47% fewer tokens per task on average.
It also reports 84.5% on BrowseComp, a test of web research by agents, and 73.3% on its FinWorkBench finance tasks. These are company-reported numbers, so treat them as a starting point and wait for independent testing before you swap anything in production.
Coding agents are where this release could change workflows fastest. If you're building with tools that plan, write and run code on their own, our guide to what agentic AI is and how agents work explains why fewer tokens per task matters so much for cost.
The security angle
Mistral spends much of its announcement on cybersecurity. The company says Large 4 scores 51% on the Artificial Analysis Cyber Index, refuses malicious cyber requests 96.6% of the time on average and resisted 97.1% of attacks on the B3 Agent Security Benchmark, up from 93.3% in preview. Mistral is also candid about risk, writing that while it saw the model refuse black-box attacks against live targets, it sees "a potential for uplift to conduct real attacks."
That candor is useful context for anyone deploying open models. Once weights are public, safety depends partly on the model's training and partly on how each team deploys it, so security reviews belong in your rollout plan.
What the license allows
Large 4 ships under what Mistral calls the Mistral Ecosystem Open-Weight license, a modified MIT license. In Mistral's words, the model is free to use, modify, fine-tune and commercialize, with one condition, which is that businesses operating a Model-as-a-Service offering must get a separate commercial license from Mistral before using it commercially.
For most startups building a product on top of the model, that reads as permissive. The gray area is for companies that sell API access to models, and AI Weekly noted that the announcement doesn't define exactly where a SaaS product ends and a Model-as-a-Service business begins. If you're anywhere near that line, read the license file on Hugging Face and talk to a lawyer before you ship.
How to try it
You can download the weights from Hugging Face and serve them with vLLM or SGLang, according to Mistral. The company also offers the model through its own API at $0.70 per million input tokens, $0.07 per million cached input tokens and $2.10 per million output tokens, and through Mistral Studio, OpenRouter and NVIDIA-hosted endpoints.
Running the full model yourself is a data center job. Mistral didn't publish hardware requirements, but in practice a trillion-parameter checkpoint needs multi-GPU servers even with quantization, so this is far from a laptop model. If local AI is your thing, our post on running Muse Spark Open on a Mac covers what a single machine can handle today. For everyday use, the API or OpenRouter is the easiest way to compare Large 4 against the assistants in our ChatGPT vs Claude vs Gemini breakdown.
Why open weights keep coming up at our events
The open versus closed question has come up again and again among Bay Area builders this year. In September we hosted the Open Weights AI Summit with Aaron Levie in Redwood City, and you can still find the event page here. It was one of the gatherings behind South Bay Renaissance, the effort Tribe coordinates to bring founders, engineers and AI researchers across the Peninsula and South Bay back together in person.
If you're testing Large 4 or comparing it with other open models, the AI Trends Tribe is a good place to share results and setups with people doing the same. And for the next AI deep dive, builders' dinner or talk, follow the Tribe Events calendar on Luma to hear about it first.
