Mistral has recently introduced the latest iteration of its open-source AI model, Mistral Large 2, promising substantial advancements in code generation, mathematics, and reasoning. Despite its smaller size compared to Meta’s Llama 3.1 405B AI model, Mistral claims that their flagship AI offers comparable performance. The Mistral Large 2 is currently available exclusively for research and non-commercial purposes.
Key Features of Mistral Large 2
Mistral announced the release of Large 2 in a detailed newsroom post, highlighting several notable features:
- Extended Context Window: The AI model boasts a context window of 128,000 tokens, matching the latest offerings from Meta.
- Multilingual Support: The model now supports an array of new languages, including Arabic, Chinese, French, German, Hindi, Italian, Japanese, Korean, Portuguese, Russian, and Spanish.
- Code Generation Capabilities: It can generate code in over 80 programming languages, a testament to its versatility.
Technical Specifications and Training
With a size of 123 billion parameters, Mistral Large 2 is designed to run on a single node, making it accessible for various research applications. The company focused on three main areas to enhance this model:
- Code Generation: The model underwent extensive training on a vast corpus of coding data, enhancing its proficiency in generating accurate and efficient code.
- Reasoning: To reduce instances of hallucination and improve logical reasoning, the model was fine-tuned to provide more cautious and reliable responses.
- Acknowledgment of Uncertainty: It is trained to recognize when it lacks sufficient information to provide a confident answer, promoting transparency and reliability.
Performance and Benchmarking
Despite being only one-third the size of the Llama 3.1 405B, Mistral claims that its Large 2 model outperforms its larger counterpart, particularly in code generation and mathematical tasks. Internal benchmarks suggest that Mistral Large 2 excels in Java code generation, even surpassing GPT-4o in this area.
Advanced Function Calling and Applications
Mistral Large 2 features enhanced function calling and retrieval capabilities, enabling it to support complex business applications. This function allows the AI to interact with external tools and databases, providing more accurate and comprehensive responses.
Availability and Access
Mistral has partnered with major cloud platforms to make Large 2 accessible:
- Google Cloud Platform: Available via Vertex AI through a managed API.
- Azure AI Studio, Amazon Bedrock, and IBM Watsonx: Accessible on these platforms as well.
For those interested in exploring this model, it is available under the name mistral-large-2407 on the company’s website. The model can also be downloaded from its HuggingFace listing under the Mistral Research Licence, which restricts usage to research and non-commercial purposes.
Final Words
The release of Mistral Large 2 marks a significant milestone in the development of open-source AI models. Its impressive capabilities in code generation, reasoning, and multilingual support, combined with its advanced function calling features, make it a powerful tool for researchers and developers alike. As it continues to evolve, Mistral Large 2 is poised to make a substantial impact on the AI landscape.