Mistral releases Large 4, a 1T-parameter open-weight MoE built for coding and cyber defense

Mistral released Mistral Large 4 on Hugging Face on October 29, under the nickname Le Chonk. The model takes text and images as input, carries roughly 1T total parameters with 49B active, and supports a context window of 1 million tokens, with official FP8 and NVFP4 checkpoints.

The license is part of the story. Mistral Large 4 ships under the Mistral Ecosystem Open-Weight License, an MIT derivative that adds restrictions on model-as-a-service offerings. The weights are public, but rehosting them commercially is not unconditional.

Architecturally, the model stacks two kinds of sparsity. Its text decoder has 72 layers, a hidden dimension of 6,144 and 64 attention heads, combining multi-head latent attention with a DeepSeek-style sparse attention indexer that selects up to 2,048 key positions per query. Three of every four decoder layers use RoPE while the fourth uses NoPE, and KV state is shared across layers.

The feed-forward side opens with a dense layer, then 71 mixture-of-experts blocks in which a softmax router picks four of 128 experts per token alongside one always-on shared expert, for five active experts per token. Vision comes from a 1.6B Pixtral encoder with 30 layers, 2,048 hidden dimensions and 16 attention heads, using 14x14 patches and a patch-merge projector with spatial merge size 2.

Cybersecurity is the standout capability. Mistral reports 82% on the Artificial Analysis Cyber Index task that requires reproducing a real open-source vulnerability and then patching it, described as the highest reported score in that index, 93% across 40 Cybench competition challenges and 93.3% attack resistance on Lakera's B3 benchmark. Internal testing also found the model useful for malware analysis, vulnerability prioritisation and writing detection rules, none of which it was explicitly trained for.

Open weights carry practical weight here. Vulnerability research and incident response often look like attacks in model input, so closed-source filters can block them, and API-dependent teams can lose access mid-incident. Mistral says the model still declines harmful requests: on cyber prompts from JailbreakBench, StrongREJECT and AgentHarm, its average refusal rate is higher than any other open-weight model.

On performance, Mistral Large 4 ranks fifth among open-weight models on the Artificial Analysis Intelligence index as of October 29. It scores 61.7% on DeepSWE v1.1, 59.4% on SWE-Atlas-QnA and 28.3% on Terminal-Bench 4, with a combined Coding Agent Index of 49.8% that places it ahead of DeepSeek V4 Pro 0813 and Qwen3.8 Max, plus 59.9% on AutomationBench and 1,393 Elo on AA-Briefcase.

Deployment friction is low by design. The model has day-0 support in transformers, vLLM and llama.cpp, GGUF quantizations from Unsloth and the llama.cpp team, remote Inference Endpoints, serverless Inference Providers, fine-tuning through TRL and deployment via Dell Enterprise Hub. For teams handling sensitive code or malware samples in their own environments, that combination matters more than a leaderboard position.