AI firm Mistral AI has launched the Mistral Large 2 model, its latest flagship AI model. The new iteration offers significant improvements on its predecessor, with considerable ability in code generation, mathematics, reasoning, and advanced multilingual support. Furthermore, Mistral Large 2 offers enhanced function-calling capabilities and is designed to be cost-efficient, high-speed, and high-performance.
Users can access the model on “la Plateforme”, with specifically developed features that make creating innovative AI applications more straightforward. Compatible languages include French, German, Spanish, Italian, Portuguese, Arabic, Hindi, Russian, Chinese, Japanese, and Korean. It also supports over 80 coding languages such as Python, Java, C, C++, JavaScript, and Bash. The model is optimized for single-node inference and long-context applications with 123 billion parameters for high throughput on a single node.
Mistral Large 2 has set a new standard for performance and cost efficiency by reaching an 84.0% accuracy score on the MMLU benchmark. Its superior performance in code generation and reasoning is due in part to the experience gained from training previous models like Codestral 22B and Codestral Mamba. It performs competitively with leading models such as GPT-4o, Claude 3 Opus, and Llama 3 405B.
The AI model’s training focused largely on enhancing its reasoning capabilities and reducing incorrect or irrelevant information generation. It also shows significant improvements in following instructions and conversational capabilities. The model is designed to give concise and cost-effective outputs which are critical for many business applications.
Mistral Large 2 distinguishes itself with its multilingual capabilities. Unlike many models that are primarily English-centric, it was trained on a considerable amount of multilingual data which enables it to excel in multiple languages, making it suitable for diverse business cases involving multilingual documents.
The model can execute parallel and sequential function calls effectively, a feature that makes it a robust engine for complex business applications. As older models are gradually phased out, all Apache models will remain available for deployment and fine-tuning, and these capabilities will be extended to Mistral Large, Mistral Nemo, and Codestral.
To conclude, Mistral AI alliances with leading cloud service providers, broadening the availability and impact of its models in the AI landscape. Mistral AI’s best models, as a result, are now accessible on Vertex AI, Azure AI Studio, Amazon Bedrock, and IBM watsonx.ai. This novel venture expands Mistral AI’s footprint to a global audience.
