Mistral Large 4 Is Now Open Weight: What You Can Run and What the License Allows

Mistral has released the weights of Mistral Large 4, its 1-trillion-parameter “Le Chonk” model, on Hugging Face, along with an improved checkpoint and a price close to half the preview’s list rate. You can download it, fine-tune it and run it on your own hardware today, but realistically that hardware is a multi-GPU server or a 512 GB Mac, not a workstation. The license is open for internal and product use, with one big exception: if you sell hosted access to the model, you need a separate commercial deal with Mistral.

Key points

• Official FP8 and NVFP4 checkpoints are on Hugging Face as mistralai/Mistral-Large-4-1T-A52B, with community GGUF quants from Unsloth and the llama.cpp team.
• It is a mixture-of-experts model with about 1T total parameters and 49B active per token, a 1M-token context window, and text and image input.
• The license is a modified MIT license. Internal use and features embedded in your own product are allowed. Running a “Model as a Service” business needs a commercial license.
• API pricing is now $0.70 per million input tokens, $0.07 cached and $2.10 output, against a $1.36 and $4.18 list price during the preview.
• Mistral reports big gains from three more weeks of reinforcement learning, including Terminal-Bench 4.0 up 13 points to 41%. Independent scores put it among the top open models, not at the top.

What did Mistral actually release?

When Mistral previewed Large 4 earlier this month, it was a paid API with weights and license listed as “coming soon.” That gap is now closed. The release includes:

• Weights: an FP8 checkpoint and an NVFP4 checkpoint for NVIDIA Blackwell GPUs, both on the Hugging Face model page. The llama.cpp team ships quants from 2-bit (IQ2_XXS, Q2_K) up to Q8_0 and BF16.
• Runtime support from day one: Transformers (v5.22.0), vLLM, SGLang and llama.cpp, plus fine-tuning through Hugging Face’s TRL library.
• Hosted options: the Mistral API as mistral-large-4, Mistral Studio, OpenRouter, Microsoft Foundry, NVIDIA-hosted endpoints and NIM containers, Hugging Face Inference Endpoints and the Dell Enterprise Hub.
• An updated model: this is not the preview checkpoint. Mistral says it kept scaling reinforcement learning and improved reasoning efficiency, so the new model uses fewer tokens for the same work.

Under the hood, Hugging Face’s technical write-up describes 72 layers, 128 experts with four routed plus one shared expert active per token, Multi-head Latent Attention combined with a DeepSeek-style sparse attention indexer, and a 1.6B-parameter Pixtral vision encoder. Reasoning has three levels: low, medium and high, with high as the default. Reasoning cannot be switched off entirely. A full technical report is still “coming soon.”

Preview vs open-weight release

What does the license allow?

This is where most of the Hacker News discussion landed. Mistral Large 3 shipped under Apache 2.0. Large 4 uses the new Mistral Ecosystem Open-Weight License, which starts from MIT and adds one condition.

• Allowed without asking: using, copying, modifying, fine-tuning, distributing and selling the model or derivatives, as long as you keep the license notice.
• Needs a commercial license: any company, or its affiliates, that operates a “Model as a Service” business and wants to use the model commercially. The license defines that as giving third parties inference or post-training access, for example via an API or hosted endpoint, where they have “meaningful control over the inputs, parameters, or training data.”
• Explicitly exempt: internal use that never exposes the model to third parties, end-user products where the model sits inside specific features or harnesses, simply relaying requests to models hosted by others, and use through Mistral’s own products.

For most builders, that means a company can run Large 4 for its own staff, or ship it inside a product feature, without calling Mistral. Inference providers and GPU clouds that want to resell tokens cannot. The gray zone is the phrase “meaningful control over the inputs.” A product that lets users send free-form prompts straight to the model looks closer to a service than a narrow feature does. If your product sits anywhere near that line, get legal advice or talk to Mistral before you launch.

Critics on Hacker News argue this makes Large 4 “weights-available” rather than open in the open-source sense, and the Open Source Initiative’s definition would agree. That is a fair point about labels. In practice you still get far more control than with any closed model: you can inspect, fine-tune and self-host it, and only one kind of business needs to sign a deal.

What hardware do you need to run it?

Only 49B parameters are active per token, which helps speed, but all of the roughly 1T parameters still have to sit in memory. These are the setups Hugging Face reported from its own runs:

The practical reading: a single 512 GB Mac can run a 2-bit quant for experiments, and a 4x professional-GPU box can serve a small team slowly. Production use with long context means a four-GPU Blackwell node with NVFP4, or an eight-GPU H200 node with FP8. Heavy quantization also costs quality, and nobody has published how much the 2-bit quants lose on this model yet. Fine-tuning is heavier still. Hugging Face’s TRL example runs LoRA on the FP8 weights across an 8x B300 node.

How good is it now?

Mistral’s own numbers for the release checkpoint are strong for an open model:

• Coding: 48% on the Artificial Analysis Coding Agent Index, effectively tied with GLM-5.3 while using 47% fewer tokens per task. It leads open models on Terminal-Bench 4.0 at 41% and on Cognition’s FrontierCode 2.0 at 19.7%.
• Human preference: in Surge’s blind coding comparisons, graders preferred it over GLM-5.3, Kimi K3 and DeepSeek V4.1 Flash. Only Claude Opus 5.5 ranked ahead of it among the models tested.
• Security: 51% on the Artificial Analysis Cyber Index, with a 96.6% average refusal rate on malicious cyber prompts, which Mistral says is the highest among open models.
• Knowledge work: 84.5% on BrowseComp for web research and 73.3% on the Finch finance spreadsheet benchmark.

Independent data is more modest. Artificial Analysis scores it 26 on its Intelligence Index v5, above the median of 23 for large open-weight models, and Hugging Face says that makes it the fifth-best open-weight model on the index. Don’t compare that 26 with the preview’s score of 38. Artificial Analysis has changed the index since then, and the scale is different. Artificial Analysis also measured about 161 output tokens per second on Mistral’s API and $2.06 per index task. It still rates the model “somewhat verbose,” using about 110M output tokens to run the index against a median of 100M.

The fair summary: Large 4 is now a credible open model for coding and security work, and the most capable from a European lab. It is not the best open model overall, and it is still behind closed frontier models such as Claude Opus 5.5.

What about the cyber guardrails?

Mistral is unusually candid here. It says that after watching live traffic since the preview, it sees “a potential for uplift to conduct real attacks,” and it is tightening guardrails over the coming months. A more capable cyber variant for penetration testing and red-teaming is going only to vetted partners, starting with Orange Cyberdefense, under contract.

For self-hosters, the refusal numbers describe default behavior, not a hard control. Anyone with the weights can prompt around or fine-tune away refusals, a point several Hacker News commenters made. If you deploy Large 4 for security work, your own access controls and logging matter more than the model’s built-in refusals.

Who should use it?

• Teams that can’t send data to US or Chinese providers: this is the clearest fit. You can self-host it, or use it through Mistral or Microsoft Foundry, and Mistral says it was trained from scratch in its own European data centers.
• Security teams: fewer refusals on legitimate defensive work, plus the option to fine-tune on your own vulnerability data, make it worth a test.
• API users chasing price: $0.70 and $2.10 is cheap for a model this size, but measure cost per finished task, since high reasoning mode uses a lot of output tokens.
• Inference providers: you need a commercial license from Mistral before serving it.
• Solo developers on one GPU: this is not your model. Mistral says Large 4 will be the base for a new generation of smaller, specialized models, which are more likely to fit your hardware.

Large 4 also joins a growing list of big open-weight releases, alongside Kimi K3.1 Pro, whose weights Moonshot has promised but not yet shipped. If you need open weights from a non-Chinese lab, Large 4 is now on the shortlist with weights you can actually download.

What to watch

• The promised technical report, and whether it explains the training data and post-training method.
• How Mistral interprets “meaningful control over the inputs” for products built on the model.
• Independent quality tests of the 2-bit and 4-bit quants that most self-hosters will actually run.
• The specialized and optimized models Mistral says it will build on Large 4.
• Any changes to the cyber guardrails Mistral says it is updating.

Key takeaways

• Mistral Large 4 weights are downloadable now in FP8 and NVFP4, with community GGUFs and day-one support in vLLM, SGLang, Transformers and llama.cpp.
• The license is MIT plus one condition: Model-as-a-Service businesses need a commercial license. Internal use and features embedded in your own product are fine.
• Self-hosting realistically starts at a 512 GB Mac for 2-bit experiments and four Blackwell GPUs for serving.
• The API dropped to $0.70 input and $2.10 output per million tokens.
• It is a top-tier open model for coding and cyber defense on Mistral’s numbers, and a solid but not leading one on independent indexes.

Sources

• Mistral: Mistral Large 4 (Le Chonk) is now open weight
• Hugging Face: Mistral-Large-4-1T-A52B model card
• Mistral Ecosystem Open-Weight License
• Hugging Face: Welcome Le Chonk, Mistral Large 4 on Hugging Face
• Artificial Analysis: Mistral Large 4 (high)
• The Verge: Mistral AI has officially released its Large 4 model
• Hacker News discussion