Mistral has officially launched Mistral Large 4, affectionately dubbed 'Le Chonk,' on Hugging Face. This new multimodal, multilingual model utilizes a Mixture-of-Experts (MoE) architecture with approximately 1 trillion total parameters and 49 billion active parameters. It accepts both text and image inputs to generate text responses and features an expansive context window of 1 million tokens. The release includes two official variants alongside GGUF quantizations created by Unsloth and the llama.cpp team, all distributed under the Mistral Ecosystem Open-Weight License, which is an MIT derivative with specific limitations on model-as-a-service offerings.
The model's architecture combines a large pool of experts with sparse attention mechanisms and a Pixtral vision encoder. The language model dynamically selects a small group of experts for each token, while an attention indexer determines which previous token positions to attend to, creating two distinct forms of sparsity. Specifically, the text decoder consists of 72 layers with a hidden dimension of 6,144 and 64 attention heads. It employs Multi-head Latent Attention (MLA) paired with a DeepSeek Sparse Attention (DSA) indexer, which selects up to 2,048 key positions per query. The first layer uses a dense feed-forward network, while the remaining 71 layers utilize MoE blocks where a router selects four of 128 experts plus one shared expert for every token.
Mistral Large 4 is positioned as a powerful tool for cybersecurity, coding agents, and multimodal applications. In independent testing by Artificial Analysis, the model scored 82% on the Cyber Index, which evaluates a model's ability to reproduce real vulnerabilities in open-source software and subsequently patch them. Mistral notes that several closed frontier models score near zero on this specific test because they refuse to attempt the reproduction phase, a step often necessary for defenders to fix issues. Beyond benchmarks, internal testing indicated strong performance in malware analysis, vulnerability prioritization, and writing detection rules, despite the model not being explicitly trained for these tasks.
Despite its offensive capabilities in security research, Mistral Large 4 maintains robust safety filters. It is primarily positioned for defensive security work, addressing the issue where closed-source models might block legitimate vulnerability research inputs or lose availability during critical incidents. On cyber prompts from JailbreakBench, StrongREJECT, and AgentHarm, the model's average refusal rate exceeds that of any other open-weight model. For the broader community, the open-weight nature allows users to build custom cyber systems around their own codebases and malware samples, fine-tune smaller models using traces from Mistral Large 4, and directly measure refusal behaviors without relying on provider descriptions.
In terms of coding performance, Mistral Large 4 achieves a Coding Agent Index score of 49.8%, placing it ahead of DeepSeek V4 Pro 0813 and Qwen3.8 Max. It scores 61.7% on DeepSWE v1.1, 59.4% on SWE-Atlas-QnA, and 28.3% on Terminal-Bench 4. Its capabilities extend to knowledge work with a score of 59.9% on AutomationBench and 1,393 Elo on AA-Briefcase. As of October 29, 2026, it ranks as the fifth-best open-weight model on the Artificial Analysis Intelligence index. Additionally, it demonstrates 93.3% attack resistance on Lakera's B3 benchmark against prompt-injection attacks.
Deployment options are extensive, with day-zero support for transformers, vLLM, and llama.cpp. The Pixtral vision encoder used in the model has 1.6 billion parameters, 30 layers, and utilizes a patch-merge projector. For local deployment, GGUF quantizations ranging from IQ2_XXS to BF16 are available, with 2-bit versions capable of running on a single Apple Silicon Ultra computer with 512 GB of unified RAM. Performance tests on an M3 Ultra showed a prefill throughput of roughly 90 tokens per second and a generation speed of 15 tokens per second for the Q2_K quantization. On four RTX6000 Pro cards, NVFP4 variants achieved prefill speeds between 130 and 590 tokens per second depending on context size.
For enterprise deployment, Mistral provides ready-to-deploy recipes on Inference Endpoints using eight NVIDIA H200 GPUs. This setup utilizes tensor and expert parallelism with the model's built-in DSpark draft head for speculative decoding, supporting up to 131K tokens of context. Under load, the system can process approximately 480 tokens per second across 1,000 ShareGPT conversations. Users can also deploy the model on Dell AI Servers using the dell-ai CLI or fine-tune it directly using TRL with support for multilingual reasoning tasks.
Sources: Hugging Face Blog.