Mistral Large 4, also called Le Chonk, is an open-weight multimodal, multilingual mixture-of-experts model released on Hugging Face. It takes text and images as input and generates text, with about 1T total parameters, 49B active parameters, and a 1M-token context window. It ships in two official variants, FP8 and NVFP4, and also has GGUFs created by Unsloth and the llama.cpp team. It is distributed under the Mistral Ecosystem Open-Weight License, an MIT derivative that places limits on model-as-a-service offerings.

The model combines a large pool of experts with sparse attention and a Pixtral vision encoder. For each token, the language model selects a small group of experts, while its attention indexer chooses which earlier token positions to attend to. The text decoder has 72 layers, a hidden dimension of 6,144, and 64 attention heads. Its attention blocks combine Multi-head Latent Attention (MLA) with a DeepSeek Sparse Attention (DSA) indexer, a direction also seen in models such as GLM 5.3; MLA keeps a compressed KV representation, and DSA selects up to 2,048 key positions per query. The decoder uses a 3:1 RoPE-to-NoPE pattern, and KV state is shared across layers. The first decoder layer is dense; the other 71 layers use MoE blocks in which a softmax router selects four of 128 experts for each token. A shared expert always runs alongside them, for five active experts per token. RoPE is LeRope, a YaRN-scaled RoPE with one learned multiplier per frequency. The vision encoder is a 1.6B Pixtral encoder with 30 layers, hidden dimension 2,048, 16 attention heads, 14×14 patches, and a patch-merge projector with spatial merge size 2.

Cybersecurity is a standout area. Mistral Large 4 scores 82% on the Artificial Analysis Cyber Index test that asks a model to reproduce a real vulnerability in open-source software and then patch it; Mistral reports that several closed frontier models score near zero on the same test because they refuse to attempt the work. This matters because fixing a vulnerability usually requires reproducing it. In Mistral’s internal testing, the model was also useful for malware analysis, vulnerability prioritisation, and writing detection rules, none of which it was explicitly trained for. The model still refuses harmful requests: on cyber prompts from JailbreakBench, StrongREJECT, and AgentHarm, its average refusal rate is higher than any other open-weight model. Mistral mainly positions the model for defensive security. Open weights put usage policy and deployment under the user’s control, which matters because vulnerability research and incident response can look like attacks in model input, and API access can be lost mid-incident. For community builders, open weights enable downstream systems built around local code bases, crash logs, and malware samples; TRL can fine-tune smaller models from Mistral Large 4 traces, and reproduce-then-patch tasks have a verifiable reward suitable for RL. Open weights also let researchers measure refusal behavior directly. The article includes an interview with Pierre Stock about the model’s cyber capabilities.

Benchmark results are strong across coding, cybersecurity, multimodal understanding, and agentic workflows. As of October 29, 2026, the model is the fifth best open-weight model on the Artificial Analysis Intelligence index. It reaches 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 score of 49.8%, ahead of DeepSeek V4 Pro 0813 and Qwen3.8 Max. On knowledge work, it scores 59.9% on AutomationBench and 1,393 Elo on AA-Briefcase. For cybersecurity, it ranks among the top five models globally on the Artificial Analysis Cyber Index, achieves 82% on vulnerability reproduction and patching (the highest reported score on that test within the index), and solves 93% of Cybench’s 40 security-competition challenges. It also shows 93.3% attack resistance on Lakera’s B3 benchmark against prompt-injection attacks.

Day-0 inference support ships for transformers, vLLM, and llama.cpp, with deployment options on Inference Endpoints and serverless Inference Providers. In Transformers, AutoModelForMultimodalLM and AutoProcessor handle text and image inputs, with reasoning_effort set to low, medium, or high; the default is high, and reasoning cannot be disabled, so “none” still uses high. A torchrun example runs the model across eight processes. vLLM supports the high-quality NVFP4 checkpoint, which has a smaller memory footprint and requires NVIDIA Blackwell GPUs, and comes with a DSpark drafter. llama.cpp and Llama App support the model; the llama.cpp team ships IQ2_XXS, Q2_K, Q4_K_M, NVFP4, Q8_0, and BF16 quants. The 2-bit quants run on a single Apple Silicon Ultra with 512 GB of unified RAM, while the others need distributed inference. On a single M3 Ultra, Q2_K gave about 90 tokens/s prefill and 15 tokens/s generation. On 4x RTX6000 Pro, NVFP4 at 32K context gave 130-590 t/s prefill and about 10 t/s generation; IQ2_XXS at short context gave 500 t/s prefill and 30 t/s generation, falling to 129 t/s prefill and 3 t/s generation at 800K context.

Fine-tuning is supported day 0 with TRL. The included example trains multilingual thinking on the FP8 variant on a node of 8xB300s, using SFTTrainer, LoRA, and DistributedConfig with FSDP and expert parallelism. Mistral Large 4 is available on Inference Endpoints with a recipe on 8x NVIDIA H200 GPUs that runs vLLM with FP8 weights split across all eight GPUs using tensor and expert parallelism, and uses a built-in DSpark draft head for speculative decoding. It serves up to 131K tokens of context and handles up to 16 requests at a time. Observed results include a time to first token of about 0.6 seconds for a 1K-token prompt, generation speed around 65 tok/s, 120-138 tok/s on chat prompts where DSpark drafts are accepted, about 4,300 tok/s total prefill for four 16K-token prompts, and about 480 tok/s total generation under load across 1,000 ShareGPT conversations. The endpoint supports reasoning_effort, tool calls, structured outputs, image inputs, and streaming. Dell Enterprise Hub also offers optimized deployment recipes through the dell-ai CLI for Dell AI Servers, for example NVFP4 on a B300 node.