French startup Mistral AI announced the release of Large 2 today, its new flagship model that is significantly more capable of code generation, mathematics, and reasoning. Mistral has also added improved multilingual support and advanced function calling capabilities with Large 2.
The Mistral Large 2 model has 123 billion parameters, allowing it to run on a single H100 node at high throughput. It supports French, German, Spanish, Italian, Portuguese, Arabic, Hindi, Russian, Chinese, Japanese, and Korean languages. In terms of coding, it supports 80+ coding languages, including Python, Java, C, C++, JavaScript, and Bash.
Mistral Large 2 is now available with open weights, but it is only available for free for research and non-commercial use. For commercial use, model licensing is required.
The performance of Mistral Large 2 (123B) is comparable to OpenAI"s GPT-4o, Claude Opus 3, and the recently released Meta"s Llama 3.1 405B on coding benchmarks. On Wild Bench, Arena Hard, and MT Bench benchmarks, Large 2 outperforms Llama 3.1 405B and Claude 3 Opus. On the popular MMLU benchmark, this new model outperforms Llama 3.1 70B and is comparable to Llama 3.1 405B.
For developers, Mistral Large 2 now comes with improved function calling and retrieval skills. It can now execute both parallel and sequential function calls, enabling developers to build complex business AI applications.
With the release of Large 2, Mistral is consolidating its model portfolio. It will have two general-purpose models, Mistral Nemo and Mistral Large, and two specialist models, Codestral and Embed. Mistral will discontinue Apache models (Mistral 7B, Mistral 8x7B and 8x22B, Codestral Mamba, Mathstral) in the future.
Microsoft and Mistral already had a partnership to make Mistral models available on Azure. Today, Mistral is expanding its partnership to Google to make Mistral models available on Google Cloud.
The consecutive releases of Mistral Large 2 and Llama 3.1 mark a significant milestone for the open AI ecosystem, providing two powerful GPT-4 level models for research and development. This rapid progress fuels the growing momentum towards a more open and collaborative AI landscape.
Source: Mistral