Welcome Le Chonk: Mistral Large 4 on Hugging Face
Mistral Large 4 (aka Le Chonk) is a 1 trillion‑parameter multimodal Mixture‑of‑Experts model with 49 B active parameters and a 1 M‑token context window that accepts text and images and outputs text. It achieves 82 % on the Artificial Analysis Cyber Index benchmark, outperforming closed‑frontier models that refuse to reproduce vulnerabilities. Engineers can run the model via Transformers, vLLM, or Llama.cpp and fine‑tune downstream models with TRL for security‑focused tasks.
⚡ Key Takeaways
- 82 % score on the Artificial Analysis Cyber Index test for reproduce‑then‑patch vulnerability tasks.
- Architecture combines sparse expert selection (Mixture‑of‑Experts) with a Pixtral vision encoder and an attention indexer for token‑level sparsity.
- Open‑weight model offers higher refusal rates on harmful cyber prompts (JailbreakBench, StrongREJECT, AgentHarm) than any other open‑weight model, balancing safety and utility.
- Inference can be served through Transformers, vLLM, or Llama.cpp libraries.
- Fine‑tuning smaller models from Mistral Large 4 traces is possible with the TRL library, using verifiable rewards for security‑related RL tasks.
🔧 Tools & Libraries
The open‑weight, high‑capacity design lets security teams build self‑hosted, controllable pipelines for vulnerability analysis and incident response without reliance on external APIs, while retaining strong safety refusals.
✅ Practical Steps
- Load a checkpoint from the Mistral Large 4 collection using the Transformers library for inference.
- Deploy the model with vLLM or Llama.cpp for scalable serving of multimodal inputs.
- Use TRL to fine‑tune a smaller Mistral model on trace data generated by Mistral Large 4 for code‑patching tasks.
Want the full story? Read the original article.
Read on Hugging Face Blog ↗