Mistral has introduced Mistral Large 4, nicknamed ‘Le Chonk’, a multimodal, multilingual, and mixture-of-experts model that takes text and images as input and produces text. It has approximately 1T parameters in total and a context window of 1M tokens, where a token refers to the unit of text processed by the model. It is distributed under the Mistral Ecosystem Open-Weight License, a license derived from MIT that sets limits for model-as-a-service offerings.
- Mistral Large 4 est un modèle multimodal à experts, doté d’environ 1T de paramètres et d’une fenêtre de contexte de 1M de tokens.
- Le modèle obtient 82% sur l’Artificial Analysis Cyber Index et vise surtout la cybersécurité défensive.
- Il refuse les requêtes dangereuses plus souvent que tout autre modèle open-weight, selon Mistral.
- Il fonctionne avec transformers, vLLM et llama.cpp, et peut être affiné avec TRL.
The model comes in two official variants, FP8 and NVFP4. GGUF format versions, intended for local execution, have been created by Unsloth and the llama.cpp team.
A Model Focused on Defensive Cybersecurity
According to Mistral, Mistral Large 4 scores 82% on the Artificial Analysis Cyber Index, a test that requires reproducing a real vulnerability in open-source software and then fixing it. Mistral states that several leading closed models score close to zero on this task because they refuse to attempt it. In internal tests, the model also proved useful for malware analysis, vulnerability prioritization, and writing detection rules, uses for which it was not specifically trained.
Mistral primarily emphasizes defensive use. According to the company, vulnerability research and incident response can resemble attacks in queries sent to a model and be blocked by filters. A team relying on an API can also lose access in the middle of a crisis. Open-weight models, on the other hand, give users control over usage policy and deployment. However, the model declines dangerous requests: on cybersecurity queries from the JailbreakBench, StrongREJECT, and AgentHarm benchmarks, its average refusal rate is reportedly higher than that of any other open-weight model.
Announced Architecture and Results
The text decoder has 72 layers, a hidden dimension of 6,144, and 64 attention heads. It combines Multi-head Latent Attention (MLA) and a DeepSeek Sparse Attention (DSA) indexer, which selects up to 2,048 key positions for each query. In the MoE layers, a router selects four experts out of 128 for each token, plus a shared expert, making five active experts per token. The image encoder is a 1.6B Pixtral encoder.
Mistral announces scores of 61.7% on DeepSWE v1.1, 59.4% on SWE-Atlas-QnA, and 28.3% on Terminal-Bench 4, for a combined Coding Agent Index of 49.8%, ahead of DeepSeek V4 Pro 0813 and Qwen3.8 Max. On Cybench, the model solves 93% of a series of 40 challenges from security competitions, and it achieves 93.3% resistance to attacks on the Lakera B3 benchmark. According to the source, as of October 29, 2026, it is the fifth-best open-weight model on the Artificial Analysis Intelligence index.
Deployment and Adaptation
Mistral Large 4 is compatible from launch with transformers, vLLM, and llama.cpp. On a server with eight NVIDIA H200 GPUs via Inference Endpoints, the FP8 variant reaches, for a single query, about 65 tokens per second, and up to 120 to 138 tokens per second on conversational prompts. On a Mac M3 Ultra, the Q2_K version shows about 15 tokens per second in generation.
The model can also undergo fine-tuning, meaning additional adjustment on specific data. An example is provided with TRL, Hugging Face’s library, on an eight-GPU B300 node, to teach the model to reason in the language requested by the system prompt. The TRL examples include a Slurm launcher. Dell also offers an optimized recipe for deploying the model on its Dell AI servers, including on an NVIDIA B300 node.
