Mistral Opens Large 4's 1T Weights but Makes Model-as-a-Service Hosts Buy a License

TL;DR

Mistral released the weights of Mistral Large 4 ("Le Chonk") on October 29, keeping the end-of-October promise it made at the public preview on October 6. The checkpoint is a 1-trillion-parameter mixture-of-experts with 49B active parameters per token, a 1M-token context, and image input. It also got about three weeks of extra reinforcement learning, which Mistral says lifted Terminal-Bench 4.0 from 28% to 41%. The license question the preview left open now has an answer: the Mistral Ecosystem Open-Weight License, which is MIT plus one clause that sends Model-as-a-Service companies to Mistral's sales team. If you self-host it for your own product, you can use it commercially for free. If you sell inference to other people, you have to talk to Mistral first.

What actually shipped

The official Hugging Face repo holds an FP8 checkpoint split across 54 safetensors shards, about 1.06 TB in total. A second official NVFP4 checkpoint comes in at about 610 GB. Mistral says it targets Blackwell GPUs and runs on vLLM and SGLang with "same performance."

From the model card:

• ~1T total, 49B active per token. The repo name says A52B because the 52B count includes embeddings and output layers. There are 128 routed experts, 4 active per token, plus 1 shared expert.
• 1M-token context, text and image in, text out. The vision encoder is Pixtral-style.
• Three reasoning efforts: low, medium, and high (the default). low replaces the preview's "None" mode.
• An embedded speculative-decoding head called DSpark: 3 dense draft layers in BF16 that propose 3 tokens at a time. Mistral measured up to 1.9x higher throughput with unchanged outputs.

DSpark is the most interesting engineering in the drop. A small draft network guesses the next three tokens cheaply, and the full trillion-parameter model checks all three in one forward pass, keeping the prefix it agrees with. Think of an intern who drafts the next three words of every sentence and a senior editor who can approve them at a glance: when the intern guesses right, the senior writes three words for the price of one. The output is identical to the big model running alone because the big model has the final say on every token.

Day-one support covers vLLM (Mistral's recommended path, with a dedicated Docker image), SGLang, Transformers, and llama.cpp. For fine-tuning, the card points to Axolotl, Unsloth, and TRL. Hugging Face published its own walkthrough the same day.

The license, clause by clause

The preview post ended on "license TBD." The answer is shorter than most model licenses. It is the MIT text with a second condition added, and that condition decides who pays.

• Granted to everyone: use, copy, modify, merge, publish, distribute, sublicense, and sell copies of the model. You can also further train it, build derivative works, and pass the same rights on.
• The condition: "If the Licensee or any of its affiliates operates a Model as a Service business," it needs a separate commercial license from Mistral "before using the Model or its derivative works for any commercial purpose."
• What counts as MaaS: giving a third party access to the model or a derivative "for inference or post-training by any means (for instance via API or hosted endpoints)" in a way that lets them "exercise meaningful control over the inputs, parameters, or training data."
• What does not count: end-user products where the model is "solely embedded within specific features or harnesses," and "mere relaying of requests to models hosted by others."
• Always exempt: internal use, meaning anything that does not expose the model to third parties, and use through Mistral's own products.

The trigger attaches to the company, not the deployment. A business that runs an inference API, or has an affiliate that does, needs a commercial license for any commercial use of Large 4, even inside a product that would otherwise be exempt. For everyone else the license works like a gym membership that is free for everybody except other gyms.

The Hacker News thread on the release passed 1,600 points and argued about labels. One commenter wrote that "the limits in the license mean it isn't open source," and others pointed to HashiCorp's move to a business-source license. For builders the consequence is narrower. Inference clouds were the obvious day-one hosts for a model this big, and the clause covers them. Mistral now controls where Large 4 runs as a paid API, and the license still lets an individual company self-host it.

What three more weeks of RL bought

The weights also differ from the preview. In its release post, Mistral says it kept scaling reinforcement learning after October 6 and reports these preview-to-release numbers. All of them are Mistral's own figures.

Against other open models, the leads Mistral claims are real but thin:

• Terminal-Bench 4.0: 41%, against 40% for GLM-5.3 on Mistral's chart.
• Cognition's private FrontierCode 2.0: 19.7%, against 19.3% for DeepSeek V4.1 Flash. Call that a photo finish.
• Artificial Analysis Coding Agent Index: 48%, "effectively tied" with GLM-5.3, while using 47% fewer tokens per task.
• Artificial Analysis Intelligence Index v5: 26, which the model card calls the highest among Western open-weight models. It also concedes that "the leading open-weight models from China score higher on this aggregate index."
• Surge Code blind pairwise preference: graders preferred Large 4 over GLM-5.3 and Kimi K3 in 64% of prompts and over DeepSeek V4.1 Flash in 56%. Against Claude Opus 5.5 it won 22%.

Token efficiency is the more useful result for anyone paying for inference. A model that ties the leader while using about half the tokens costs about half as much per finished task at the same per-token price.

The cyber caveat Mistral wrote itself

Mistral markets Large 4 to defenders. It reports 51% on the Artificial Analysis Cyber Index and a 96.6% average refusal rate on malicious cyber prompts, the highest it measured among open models. The release post also includes an unusually direct admission. While monitoring live preview traffic, Mistral "observed the model refusing to comply with black-box attacks against live targets," but it says it sees "a potential for uplift to conduct real attacks" and is changing its guardrails.

A specialized variant tuned for penetration testing and red-teaming exists, but only enterprise design partners such as Orange Cyberdefense get it, after vetting and under contract. The general checkpoint is the one with refusals built in. Once the weights are public, anyone can fine-tune those refusals away, and Mistral's guardrail updates will reach its API but not the copies already downloaded.

Can you actually run it?

For most homelabs, not really. The official vLLM command uses tensor parallelism across 8 GPUs. Community GGUFs from Unsloth and the llama.cpp team cover the range from desperate to lossless, and none of them is small.

With 49B active parameters, each token only touches a slice of the weights, so CPU-RAM offload with llama.cpp is slow rather than impossible. You still need roughly a quarter to two-thirds of a terabyte of memory somewhere. The realistic self-hosting audience is companies with an 8-GPU Blackwell node and a data-residency reason to own the model, which is the buyer that the bank, telecom, and security-vendor quotes in the release post are aimed at.

Everyone else can use the API. From October 29, Mistral lists Large 4 at $0.70 per million input tokens, $0.07 cached, and $2.10 output. The preview's list price was $1.36/$4.18 with a temporary half-price launch rate, so the launch discount has effectively become the regular price.

Key Takeaways

• The weights are real and downloadable: about 1T total and 49B active parameters, 1M context, image input, published October 29 in FP8 (about 1.06 TB) and NVFP4 (about 610 GB).
• The license is MIT plus one clause: a company that runs a Model-as-a-Service business, directly or through an affiliate, needs a separate commercial license, and internal use stays exempt.
• The release beats the preview: Mistral reports Terminal-Bench 4.0 up from 28% to 41% with 30% fewer output tokens, and gains on Finch and B3.
• Open-model leads are narrow: 0.4 points on FrontierCode 2.0 and a tie on the Coding Agent Index, though it gets there with 47% fewer tokens. Chinese open models still lead the aggregate intelligence index.
• Mistral flags its own cyber risk: it sees "a potential for uplift" for real attacks and keeps a pentest-tuned variant behind vetting.
• It does not fit a homelab: the smallest GGUF is 233 GB, and the API at $0.70/$2.10 per million tokens is the practical route for most builders.

Sources: Mistral: Mistral Large 4 (Le Chonk) is now open weight, Hugging Face model card, Mistral Ecosystem Open-Weight License, NVFP4 checkpoint, Hugging Face: Welcome Le Chonk, Unsloth GGUFs, ggml-org GGUFs, Mistral docs: Mistral Large 4, Hacker News discussion