Mistral Large 4 Arrives on Hugging Face: A 1‑Trillion‑Parameter MoE Model for Cybersecurity and Multimodal Tasks

Mistral has published Mistral Large 4, nicknamed “Le Chonk,” on the Hugging Face hub. The model weighs roughly one trillion total parameters with 49 billion active at inference time and offers a one‑million‑token context window.

Its architecture combines a sparse mixture‑of‑experts language core with a Pixtral vision encoder. For each token the router activates a small subset of experts, while a separate attention indexer selects which prior positions to attend to, delivering two independent sparsity mechanisms.

Benchmark results highlight a strong cybersecurity focus. The model scores 82 percent on the Artificial Analysis Cyber Index, reproducing and patching real vulnerabilities where many closed models refuse to engage. Internal tests also show usefulness for malware analysis, vulnerability prioritisation and detection‑rule generation.

Why it matters for GPU and AI infrastructure

Running a 1 trillion‑parameter MoE model demands high‑memory GPUs and efficient sparse‑kernel support. The FP8 and NVFP4 quantised checkpoints reduce memory pressure, while the 1 M token context stresses KV‑cache capacity. Teams deploying on premises can fine‑tune smaller distilled models from Large 4 traces using TRL, turning the massive checkpoint into a practical asset for custom security pipelines.