Mistral released the weights of Mistral Large 4, nicknamed Le Chonk, on October 29, 2026. Its launch post says: "Today, we're releasing the weights of this improved Mistral Large 4." The model went into public preview on October 6, 2026, when Mistral's preview post promised: "Weights drop end of this month." Mistral says three weeks of extra reinforcement learning since then improved its coding scores. The weights are on Hugging Face under a modified MIT license with one commercial condition.

#How big is Mistral Large 4?

Mistral describes it as "our flagship 1-trillion-parameter, natively multimodal model." Its own pages give different active-parameter figures:

  • Model card. The Hugging Face model card says it has "49 billion active parameters per token (52 billion including embeddings and output layers)." The same card's feature list says "~1T total / ~52B active parameters."
  • October 6, 2026 post. "ML4 is a 1 trillion-parameter natively multimodal model with 52 billion active parameters."
  • Docs. Mistral's docs page says "It features 52B active parameters and 1.05T total parameters, and a 1.6B vision encoder."

So the 49 billion and 52 billion figures count different things, per the model card, and the total is "1 trillion" in Mistral's posts but "1.05T" in its docs.

The card lists "128 routed experts (4 active per token) and 1 shared expert." It "supports a 1M context window" and "accepts text and image input, with text output." Mistral says it is "natively fluent in 160+ languages, including every official language of the European Union," and was "trained from scratch on 3,800 NVIDIA Grace Blackwell GPUs in Mistral's own datacenters in Europe."

Mistral says the model "now supports three reasoning modes," with High as the default. The Hugging Face blog adds: "note that reasoning cannot be disabled."

#What does the license allow?

The launch post says the model uses the Mistral Ecosystem Open-Weight License, "a modified MIT license." Its plain-language version of the condition reads:

"The model is free to use, modify, fine-tune, and commercialize, with a single condition: businesses that operate a Model-as-a-Service offering must obtain a separate commercial license from Mistral before using the model commercially."

Clause 2 of the LICENSE file reads:

"If the Licensee or any of its affiliates operates a Model as a Service business, the Licensee must enter into a separate commercial license with Mistral before using the Model or its derivative works for any commercial purpose. The Licensee may contact Mistral, using this form: https://mistral.ai/contact/, to request a commercial license."

The file defines the term: "Model as a Service" means "giving a third party access to the Model or its derivative works for inference or post-training by any means (for instance via API or hosted endpoints), in a manner that allows such third party to exercise meaningful control over the inputs, parameters, or training data." It adds: "This does not include (a) end-user products with model capabilities solely embedded within specific features or harnesses, or (b) mere relaying of requests to models hosted by others."

Clause 3 exempts "internal use of the Model or its derivative works" and "any use of the Model accessed through Mistral’s official products." As of October 29, 2026, the Hugging Face repository is public and not gated.

#How does it perform?

All figures here are Mistral's; we have not tested the model. Mistral says it "ranks as the best open-weight model for coding and among the world's most capable open models overall."

  • FrontierCode 2.0. According to Mistral, it "leads with a score of 19.7%" on Cognition's private benchmark.
  • Coding Agent Index. Mistral says it scores 48% on the Artificial Analysis index, "effectively tied with the leading open-weight model, GLM-5.3, while using 47% fewer tokens per task on average."
  • Terminal-Bench 4.0. Mistral says it "improved by 13 percentage points to 41% compared to its public preview."
  • Human preference. In Surge Code's evaluations, Mistral says it ranked "behind only Claude Opus 5.5 among the models evaluated," ahead of GLM-5.3, Kimi K3 and DeepSeek V4.1 Flash.
  • Cybersecurity. Mistral says it scores 51% on the Artificial Analysis Cyber Index and has "an average refusal rate of 96.6% on malicious cyber requests, the highest we measure among open models."
  • Other tasks. Mistral reports 73.3% on Finch, 84.5% on BrowseComp, and resistance to "97.1% of attacks vs 93.3% in Preview" on the B3 Agent Security Benchmark.

The model card adds a caveat to its 26 score on the Artificial Analysis Intelligence Index v5: "The leading open-weight models from China score higher on this aggregate index, but Mistral Large 4 is particularly strong on coding and cybersecurity."

For comparison, see our coverage of DeepSeek V4.1 Flash 1017 and Recursive's Flint, another open-weight coding model with a named license.

#What about cybersecurity risks?

Mistral is direct about the risk. Its launch post says: "Though we observed the model refusing to comply with black-box attacks against live targets, we see a potential for uplift to conduct real attacks. We are modifying our guardrails accordingly, and will continue updating them in the months to come."

Mistral also gives enterprise design partners such as Orange Cyberdefense "preview access to a specialized variant of Mistral Large 4 optimized for advanced cybersecurity workflows." It says "Access is subject to rigorous vetting and contractual safeguards."

WIRED quoted Rafe Pilling, senior director of threat intelligence at Sophos: "It’s clearly a dual-use technology." WIRED also quoted Pierre Stock, Mistral's VP of science: "It’s a pivotal moment in the global AI leadership race."

#Where can you get it, and what does it cost?

Mistral says: "For full control, download the open weights from Hugging Face and deploy on your own infrastructure with vLLM or SGLang." The card also mentions "A quantized NVFP4 checkpoint." Its serving examples use "a node of 8 GPUs"; Mistral states no minimum.

Hosted options, per Mistral, are Mistral Studio, the Mistral API (as mistral-large-4) or OpenRouter, plus Microsoft Foundry and NVIDIA-hosted endpoints. Mistral's API price as of October 29, 2026 is $0.70 per million input tokens, $0.07 per million cached input tokens and $2.10 per million output tokens.

Our analysis: A trillion-parameter open-weight model from a European lab gives companies a large alternative to Chinese open-weight models, and Mistral's API price is low for its size. Two things need care. The license is not unconditional: anyone running a hosted service on the model should read clause 2 before building on it. And Mistral's own warning about "a potential for uplift to conduct real attacks" means its guardrails, not just its benchmark scores, deserve scrutiny.

#What is still unclear?

  • Self-hosting hardware. Mistral gives no minimum.
  • The cyber variant. Mistral does not say whether the specialized variant will be released more widely.
  • Independent results. No third party has published tests of the released model in our sources.