one year on
Mistral releases Large 2, claiming parity with frontier models at a fraction of the size
The 123-billion-parameter model arrives one day after Meta's Llama 3.1 405B and matches or beats it on code and math benchmarks while using a third of the parameters.
A day after Meta unleashed Llama 3.1 405B, Mistral AI fires back. The Paris-based startup today released Mistral Large 2, a 123-billion-parameter dense model with a 128,000-token context window, available under the Mistral Research License for non-commercial use.
Mistral says Large 2 performs on par with GPT-4o and Claude 3 Opus, and TechCrunch reports it appears to outpace Llama 3.1 405B on code generation and math, despite weighing in at 123 billion parameters. On MMLU, the pretrained version scores 84.0%. The company says it focused on reducing hallucinations by training the model to admit when it lacks information rather than fabricate answers. Large 2 also supports over a dozen languages and 80+ coding languages.
The timing is striking: two frontier-class open-weight models in two days. Mistral’s model is not truly open source—commercial self-deployment requires a paid license—but it lands on a week when the competitive landscape suddenly looks crowded.
The record
One year later — open only if you can handle spoilers
Mistral Large 2 established a benchmark for efficient frontier models, though it was quickly overshadowed by the broader open-weight trend. By mid-2026, Mistral had shifted focus toward smaller specialist models and enterprise deployments.