AI Agent Hub
Back to models
Mistral: Mistral Large 4 logo

Mistral: Mistral Large 4

Multimodal mistralai Released 2026-10-06
-- 1050.0B params 1M context Proprietary

About this model

Mistral Large 4 (nicknamed "Le Chonk") is Mistral AI's flagship open-weight, natively multimodal mixture-of-experts model, announced October 6, 2026 as a public API preview with full weights planned for release by the end of October 2026. It scales to roughly 1.05 trillion total parameters with about 52 billion active parameters per token (49 billion plus embedding and output layers), a 1.6 billion-parameter vision encoder, and up to one million tokens of context on Mistral's official model card. The model was trained from scratch on 3,800 NVIDIA Grace Blackwell GPUs in Mistral's European datacenters and is positioned for AI sovereignty, enterprise deployment, and self-hosted use once weights are available.

Mistral emphasizes frontier performance on agentic coding, long-horizon tool use, cybersecurity, and knowledge work rather than classic academic multiple-choice suites at launch. Official reporting highlights strong software-engineering agent scores (including 61.7% on DeepSWE v1.1), 59.9% on AutomationBench across hundreds of business workflows, 82% on CyberGym-E2E, and leading open-weight results on scientific coding benchmarks such as SciCode-Verified. Independent evaluators from Artificial Analysis report competitive multimodal and reasoning signals, including 76.4% on MMMU-Pro and 35.0% on Humanity's Last Exam, while LiveBench lists a 72.8% high-configuration result. The preview supports structured outputs, function calling, document Q&A, batching, and agents via the mistral-large-4 API identifier.

The model targets professionals in finance, law, engineering, and security who need multimodal document understanding, visual grounding on charts and technical drawings, and autonomous agents that can use tools without provider-level refusals blocking legitimate security research. Mistral notes ongoing reinforcement-learning post-training with substantial headroom, additional architecture details and benchmarks promised alongside the weight release, and specialized derivatives built on this foundation for vertical industries.

Benchmark Scores

HLE
35.0
DeepSWE
61.7
CyberGym
82.0
MMMU-Pro
76.4
LiveBench
72.8
AutomationBench
59.9

Technical Specs

  • Parameters: 1050.0B
  • Architecture: Mixture-of-Experts
  • Context Window: 1,000,000 tokens
  • Input Modalities: text, image

Hardware Requirements

  • Compute: API only

Pricing

Input Output Currency
0.68 / 1M tokens 2.09 / 1M tokens USD