Welcome Le Chonk: Mistral Large 4 on Hugging Face
Hugging Face introduces Mistral Large 4, an open-weight multimodal model with 1T total and 49B active parameters, optimized for cybersecurity and coding tasks, now available on its platform.
Mistral Large 4, also known as Le Chonk, is a new open-weight model from Mistral that processes both text and images to generate text responses. The model features approximately 1 trillion total parameters with 49 billion active parameters and supports a context window of up to 1 million tokens. It is distributed under the Mistral Ecosystem Open-Weight License, which restricts commercial use as a service. The model excels in cybersecurity, coding agents, and multimodal tasks, combining a large pool of experts with sparse attention mechanisms and a Pixtral vision encoder.
The model demonstrates strong performance in cybersecurity, scoring 82% on the Artificial Analysis Cyber Index test by reproducing and patching real vulnerabilities in open-source software. Mistral reports that several closed frontier models score near zero on the same test due to refusal to engage with such tasks. In internal testing, Mistral Large 4 also proved effective for malware analysis, vulnerability prioritization, and writing detection rules without explicit training for these tasks. The model is positioned primarily for defensive security work, offering users full control over deployment and usage policies.
Mistral Large 4’s architecture includes a text decoder with 72 layers, a hidden dimension of 6,144, and 64 attention heads, utilizing Multi-head Latent Attention and DeepSeek Sparse Attention. The model employs a 1.6B Pixtral encoder with 30 layers and supports deployment via Transformers, vLLM, and llama.cpp. GGUF quantized versions are available from Unsloth and the llama.cpp team. The model supports varying reasoning efforts, with a default setting of high, and includes optimizations for speed and memory efficiency.
The model delivers competitive benchmark results across coding, cybersecurity, multimodal understanding, and agentic workflows. As of the specified date, it ranks as the fifth best open-weight model on the Artificial Analysis Intelligence index. On coding benchmarks, it achieves scores such as 61.7% on DeepSWE v1.1 and 59.4% on SWE-Atlas-QnA, while cybersecurity tests show 93.3% attack resistance on Lakera’s B3 benchmark. Mistral Large 4 is available for deployment through Hugging Face’s Inference Endpoints and serverless options, with day-0 support for Transformers, vLLM, and llama.cpp.