Mistral releases open-weight Large 4 model with strong vulnerability-repair results
Hugging Face announced Mistral Large 4, a roughly 1T-parameter multimodal model under an open-weight license, highlighting an 82% score on a vulnerability reproduce-and-patch test.
In plain words
Mistral has released a very large AI model that can read text and images and is free to download under an open-weight license, meaning anyone can run it on their own systems. Its headline feature is security work: on a test where a model must recreate a real software vulnerability and then fix it, Mistral reports it scored 82%, which the company says beats closed models that often refuse the task.
What was announced
The model, nicknamed Le Chonk, accepts text and image inputs and generates text, with about 1 trillion total parameters, 49 billion active per token, and a 1 million-token context window. It uses a Mixture-of-Experts design in which a router selects a small group of expert networks for each token, combined with sparse attention that picks which earlier positions to review. It ships in FP8 and NVFP4 variants under the Mistral Ecosystem Open-Weight License, an MIT derivative with limits on model-as-a-service use, and community GGUF conversions exist for local running.
On security benchmarks, Mistral reports 82% on the Artificial Analysis Cyber Index reproduce-and-patch task, described as the highest reported score, and 93% on a 40-challenge Cybench suite. The company positions the model for defensive work, arguing that open weights let security teams control deployment and policy during incidents, since vulnerability research can look like attacks to API-side filters. The model still refuses harmful requests at a rate Mistral says is higher than any other open-weight model on three cyber prompt benchmarks. Benchmarks come from Mistral's announcement, not independent testing.
What it does not show
- The 82% cyber score, refusal rates, and benchmark figures come from Mistral's own announcement; the post cites Artificial Analysis as the index source but the results are not independently confirmed here.
- The source is a Hugging Face blog post about the release; it does not provide the underlying benchmark papers or full evaluation methodology.
- Hardware requirements are substantial: the NVFP4 checkpoint needs NVIDIA Blackwell GPUs, and 2-bit GGUF quants need 512 GB of unified RAM, with other quants requiring distributed inference.
For security and IT teams
- Scope
- The model is positioned for defensive security tasks such as vulnerability reproduction, patching, malware analysis, and detection-rule writing, with open weights keeping usage policy under the deployer's control.
- Access
- Weights are available in FP8 and NVFP4 variants plus community GGUFs, with day-0 support in transformers, vLLM, and llama.cpp, and deployment recipes for Inference Endpoints on 8x H200 GPUs.
- Status
- The license is the Mistral Ecosystem Open-Weight License, an MIT derivative that imposes limitations on model-as-a-service offerings.
Sources
Source-support review completed 14 Nov 2026. This story was drafted with AI assistance and checked against retained source passages. Automated checking is not independent verification.
Updates and corrections
No corrections so far.
First published 14 Nov 2026
New to the subject? AI from the Ground Up starts from the beginning.