Welcome Le Chonk: Mistral Large 4 on Hugging Face
Web page · https://huggingface.co/blog/mistral-large4
First seen by Oriel on .
AI summary
The post introduces Mistral Large 4, an open-weight multimodal MoE model with a 1M-token context window and reported strengths in cybersecurity, coding agents, and visual understanding.
- It has about 1T total parameters, 49B active parameters, a 1M-token context window, FP8 and NVFP4 variants, and an Mistral Ecosystem Open-Weight License.
- Its decoder has 72 layers; 71 use MoE blocks routing four of 128 experts plus one shared expert, while DSA selects up to 2,048 key positions.
- It scores 82% on Artificial Analysis Cyber Index vulnerability reproduction-and-patching, solves 93% of 40 Cybench challenges, and reports 93.3% attack resistance on Lakera’s B3.
- Coding results include 61.7% on DeepSWE v1.1, 59.4% on SWE-Atlas-QnA, and a 49.8% Coding Agent Index score.
- Transformers, vLLM, and llama.cpp support the model; NVFP4 vLLM requires NVIDIA Blackwell GPUs, while 2-bit GGUF quants run on one Apple Silicon Ultra with 512 GB unified RAM.
#mistral #mixture of experts #multimodal models #cybersecurity
Generated by gpt-5.6-sol. Check the original for details.
Aliases
| Source | External ID | URL |
|---|---|---|
| Hugging Face Blog | https://huggingface.co/blog/mistral-large4 | https://huggingface.co/blog/mistral-large4 Read discussion |