Mistral Large 2: 123B Open-Weights Model That Matches GPT-4 on Code and Logic

Mistral released the 123B parameter Large 2 model with downloadable weights and an API endpoint. On coding and reasoning benchmarks the model scores within a few points of GPT-4 while supporting 32k context. Users access it through the Mistral API or by running the weights with vLLM or Hugging Face Text Generation Inference.

Comparable performance at lower per-token cost removes the need to default to closed frontier models. Developers can host the model on their own infrastructure for compliance or cost control. The release normalizes treating high-end models as interchangeable components rather than vendor commitments.

Mistral reports enterprise pilots where teams replaced GPT-4 calls with Large 2 hosted on AWS g5 instances, cutting inference spend by roughly 40 percent while maintaining pass@1 scores above 75 percent on HumanEval.