Mistral AI released the weights of Mistral Large 4 on October 29, 2026, making its largest model, nicknamed “Le Chonk,” free to download for companies and governments. The model has about one trillion parameters, takes text and images as input and handles a 1-million-token context window. Mistral is pitching it first at cybersecurity teams that want to run a capable model under their own policies, rather than depend on a closed provider’s filters.
What was released
Mistral first showed Large 4 on October 6 as a public preview through its API and promised weights by the end of the month. Hugging Face, which published a launch walkthrough on October 29, describes a multimodal Mixture-of-Experts model with around 1 trillion total parameters and a 1M-token context window. Mistral’s own announcement says 52 billion parameters are active per token, while the Hugging Face post says 49 billion, so we cite both figures as published rather than choose between them.
Official checkpoints come in FP8 and NVFP4 variants, and the llama.cpp team and Unsloth have produced GGUF files. The license is the Mistral Ecosystem Open-Weight License, which Hugging Face describes as an MIT derivative that adds limits on offering the model as a service. That makes Large 4 open-weight rather than fully open source in the usual sense. See our explainer on open-weight versus closed models for why the distinction matters in practice.
How it is built
According to Hugging Face, the text decoder has 72 layers. The first layer is dense and the other 71 are Mixture-of-Experts blocks in which a router picks four of 128 experts per token, plus one shared expert that always runs. Attention combines Multi-head Latent Attention with a DeepSeek Sparse Attention indexer that selects up to 2,048 key positions per query, a design aimed at keeping long-context inference affordable. A 1.6-billion-parameter Pixtral encoder handles images. Mistral says the model was trained from scratch on 3,800 NVIDIA Grace Blackwell GPUs in its own European data centers, with training data spanning more than 160 languages including every official EU language.
The cybersecurity pitch
Mistral argues that closed models often refuse legitimate security work, because vulnerability research and incident response look like attacks in a prompt, and that a team relying on an API can lose access mid-incident. Open weights, it says, put the usage policy and the deployment in the customer’s hands. Mistral claims 82% on the Artificial Analysis Cyber Index task that asks a model to reproduce a real open-source vulnerability and then patch it, and Hugging Face reports 93% on Cybench, a set of 40 security-competition challenges, and 93.3% attack resistance on Lakera’s B3 benchmark. The company also says the model’s refusal rate on harmful cyber prompts from JailbreakBench, StrongREJECT and AgentHarm is higher than any other open-weight model.
TechCrunch reports that Mistral considers Large 4 comparable to China’s GLM-5.3 on cyber work while using 47% fewer tokens per task. Mistral’s figures are its own; only the Artificial Analysis placement is an independent measure.
Where it ranks
Hugging Face says that as of October 29 the model is the fifth-best open-weight model on the Artificial Analysis Intelligence Index. On coding it lists 61.7% on DeepSWE v1.1, 59.4% on SWE-Atlas-QnA and 28.3% on Terminal-Bench 4, and a Coding Agent Index score of 49.8%, ahead of DeepSeek V4 Pro 0813 and Qwen3.8 Max. The Decoder’s headline framing is more cautious: a strong coding model that still trails Chinese labs overall. Wired reported Mistral’s own claim that Le Chonk has leapfrogged China in the open-weight race, which the index placement does not support.
Why it matters
Open weights let governments and regulated firms run a near-frontier model on their own hardware, which is the core argument in our explainer on sovereign AI. It also lets researchers measure refusal behavior directly instead of relying on a vendor’s description. The release lands a day after Google limited Gemini 4 Carbon to a small group of defenders, as we report in our Gemini 4 Carbon story. Two labs with opposite distribution models are now both targeting security work. For more on the company, see our Mistral AI profile.
What to watch next
- Independent evaluations of the model’s cyber results, including whether the 82% figure holds outside Mistral’s setup.
- Fine-tunes and smaller models built from Large 4 traces, which Hugging Face expects the community to try.
- Whether misuse reports appear, given that anyone can now remove guardrails from downloaded weights.
- How governments and large enterprises respond to the European deployment option Mistral says it operates end to end.
Key takeaways
- Mistral Large 4 weights are now downloadable under the Mistral Ecosystem Open-Weight License.
- The model has roughly 1T total parameters, a 1M-token context window and text-and-image input.
- Mistral and Hugging Face report 52B and 49B active parameters respectively.
- Cyber results (82% reproduce-and-patch, 93% Cybench) come from Mistral and Hugging Face; the index ranking is fifth among open-weight models.
- The release strengthens the open-weight, sovereign-AI argument against closed cyber models.
Frequently asked questions
What is Mistral Large 4?
Mistral Large 4, nicknamed Le Chonk, is a multimodal Mixture-of-Experts model with about 1 trillion parameters and a 1M-token context window. Mistral released its weights on October 29, 2026.
Is Mistral Large 4 open source?
It is open-weight. The Mistral Ecosystem Open-Weight License is an MIT derivative that restricts offering the model as a service, so it is not fully open source.
Why does Mistral focus on cybersecurity?
Mistral says defenders need models that will not refuse vulnerability research and that stay available during incidents. Self-hosting open weights gives them that control.
