Mistral Large 4 gains ground in code and cybersecurity
The Mistral Large 4 model, nicknamed “Le Chonk,” arrives on Hugging Face with text, images, and a context of one million tokens. Mistral highlights it for code, cybersecurity, and agent use cases, with open weights under a restrictive license.
Mistral Large 4, nicknamed "Le Chonk," is a multimodal and multilingual model from Mistral now available on Hugging Face. Capable of processing text and images to generate text, it is based on a Mixture-of-Experts architecture, with around 1 trillion parameters in total, 49 billion activated, and a context window of one million tokens. Its arrival is especially significant for teams wanting to deploy, fine-tune, or audit a large model without relying on a single API, particularly in cybersecurity and for code agents.
An open model, but not completely free
Mistral distributes Mistral Large 4 under the Mistral Ecosystem Open-Weight License, a variant inspired by Apache/MIT that imposes limits on model-as-a-service offerings. In short, the weights are accessible, but commercial use is not a copy-paste without conditions. Two official variants are offered, FP8 and NVFP4, along with GGUF versions created by Unsloth and the llama.cpp team.
This positioning is strategic. In practice, an open weight model offers more control over deployment, filters, and behavior auditing. Mistral also emphasizes a point rarely highlighted so clearly: an open model allows verification of refusals, not just taking them at face value. A small phrase that stings a bit, but clearly states the challenge.
Why cybersecurity is at the heart of the message
Mistral presents Large 4 as particularly strong for cybersecurity, code, and multimodal uses. On the Artificial Analysis Cyber Index, it scores 82% on a task of reproducing then fixing a real vulnerability in open-source software. The company explains that several closed models score close to zero because they often refuse to engage with this type of request.
The issue goes beyond the benchmark score. For a security team, reproducing a flaw before fixing it is often essential. Yet a system dependent on an API can block at the worst moment, when the incident is already underway. Mistral thus highlights a model that remains deployable locally or at a third party, while maintaining a high refusal rate on malicious requests according to JailbreakBench, StrongREJECT, and AgentHarm [to be verified on the exact scope of the measures].
An architecture designed to be fast without doing everything at once
Technically, Mistral Large 4 combines several forms of efficiency. The text decoder has 72 layers, a hidden dimension of 6,144, and 64 attention heads. Its attention combines Multi-head Latent Attention and a DeepSeek Sparse Attention indexer, which selects up to 2,048 past positions per query. Another detail: the model alternates RoPE and NoPE in a 3:1 pattern, with KV states shared between layers.
The MoE structure does the rest. After the first dense layer, the following 71 layers rely on 128 experts, of which four are selected per token, plus one shared expert. This allows increasing capacity without activating the entire network for each query. For images, Mistral adds a Pixtral encoder with 1.6 billion parameters, 30 layers, and 14 × 14 patches.
Scores that also target the agent field
The published results place it among the top open weight models, although model comparisons remain delicate. As of October 29, 2026, Mistral states that Large 4 ranks as the fifth best open weight model on the Artificial Analysis Intelligence Index. On code, it scores 61.7% on DeepSWE v1.1, 59.4% on SWE-Atlas-QnA, and 28.3% on Terminal-Bench 4. Its combined score of 49.8% on the Coding Agent Index places it ahead of DeepSeek V4 Pro 0813 and Qwen3.8 Max, according to Mistral's source.
The model is not limited to software development. It also scores 59.9% on AutomationBench and 1,393 Elo on AA-Briefcase, two evaluations related to business workflows and long professional tasks. This is precisely the kind of profile that attracts publishers wanting a model capable of reading, deciding, and acting, without promising the moon. In AI, that’s already quite something.
Local, cloud, or enterprise server deployment
Mistral Large 4 comes with support from day one for Transformers, vLLM, and llama.cpp. Hugging Face also announces ready-to-use Inference Endpoints on 8 NVIDIA H200 GPUs, with up to 131,000 tokens of context and 16 parallel requests. On performance, the platform reports about 0.6 seconds to the first token for a 1,000-token request, then up to 65 tokens per second in generation, with higher peaks on some chat prompts.
Teams preferring on-premise can use llama.cpp, which already supports GGUF quants, or Dell Enterprise Hub on Dell AI servers. The NVFP4 checkpoint requires NVIDIA Blackwell GPUs, while some lighter versions can run on more modest setups. In short, everyone can choose their hardware luxury level, from research studio to fully loaded rack.
What this release changes for the market
This release confirms a clear trend: open weight models are no longer just for research or prototyping. They also aim to take place in sensitive uses, from code to security to business automation. With Large 4, Mistral targets a segment where technical sovereignty, fine-tuning capability, and cost control matter as much as raw model quality.
One structural limit remains: access to weights does not mean absence of constraints. The license governs commercial exploitation, and the best results require serious, sometimes costly infrastructure. But for companies wanting to keep control over their models, Mistral Large 4 becomes a hard-to-ignore option.
Key points
- Released on Hugging Face on October 29, 2026.
- About 1 trillion parameters, 49 billion activated.
- Context window of one million tokens.
- 82% on the reproduced then fixed vulnerability test.
- Four experts out of 128 activated per token.
- Hugging Face, llama.cpp, and vLLM supported from launch.
By the numbers
- 1 trillion — total model parameters, source Mistral/Hugging Face.
- 49 billion — active parameters, source Mistral/Hugging Face.
- 1 million tokens — announced context window, source Mistral/Hugging Face.
- 82% — score on the cyber test of reproduction then correction, source Mistral.
- 1.6 billion — parameters of the Pixtral visual encoder, source Mistral.