NeuralStackMistral AI launched Large 3 on July 21 2026 with a 4 million token context window and native support for 32 languages. The model achieves 92.4 percent on MMLU benchmarks and 89.1 percent on HumanEval coding tasks. It processes inputs at 180 tokens per second on H100 clusters.
Training completed on July 15 using 18,000 H100 GPUs over 42 days with a final dataset of 14 trillion tokens. Mistral integrated reinforcement learning from human feedback updates focused on safety and factual accuracy. The company opened API access immediately with rate limits starting at 10,000 requests per minute.
Background includes Mistral's prior Large 2 release in March 2025 that reached 2 million tokens. European regulators have pushed for transparent model cards since the 2025 AI Act updates. Large 3 training data includes licensed European news and academic corpora to address prior copyright concerns.
Why this matters
European companies gain a high-performance alternative that complies more easily with upcoming audit requirements. The 4 million token window enables new use cases in legal document analysis and long-form codebases. Mistral's open weights policy for smaller variants may accelerate adoption among mid-sized developers.
Industry analysts expect Large 3 to capture 12 percent of the European LLM API market by December 2026. Forward pricing starts at 8 dollars per million input tokens and 24 dollars for output. The release pressures US labs to accelerate their own context length improvements before year end.