Mistral AI has introduced Mistral Large 2, the latest version of its flagship language model, which offers significant improvements in code generation, mathematics, reasoning, multilingual support, and function calling capabilities.

Key Features

• Context Window and Language Support: Mistral Large 2 boasts a 128k context window and supports multiple languages including 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.
• Single-Node Inference: Designed for single-node inference, it has 123 billion parameters, enabling high throughput on a single node.
• Licensing: Available under the Mistral Research License for research and non-commercial use, with a commercial license available upon request.

Performance

• General Performance: Achieves 84.0% accuracy on MMLU, setting a new benchmark for performance/cost efficiency.
• Code & Reasoning: Trained extensively on code, it matches leading models like GPT-4o and Llama 3 405B. Efforts were made to reduce "hallucinations" and improve accuracy in responses.
• Mathematical Benchmarks: Shows enhanced performance on mathematical benchmarks like GSM8K and MATH.

Instruction Following & Alignment

• Improved Capabilities: Enhanced in following precise instructions and handling long multi-turn conversations. Performance is measured on benchmarks like MT-Bench, Wild Bench, and Arena Hard.
• Conciseness: Focused on generating succinct responses to facilitate quicker interactions and cost-effective inference.

Language Diversity

• Multilingual Proficiency: Trained on a large proportion of multilingual data, excelling in languages such as English, French, German, Spanish, Italian, Portuguese, Dutch, Russian, Chinese, Japanese, Korean, Arabic, and Hindi.
• Benchmark Performance: Outperforms previous models on the multilingual MMLU benchmark.

Tool Use & Function Calling

• Enhanced Functionality: Equipped with advanced function calling and retrieval skills, capable of executing both parallel and sequential function calls for complex business applications.

Availability

• La Plateforme: Available under the name mistral-large-2407 on la Plateforme and can be tested on le Chat. The model weights are also hosted on HuggingFace.
• Cloud Service Providers: Available on Google Cloud Platform's Vertex AI, Azure AI Studio, Amazon Bedrock, and IBM watsonx.ai.

Mistral Large 2 represents a significant leap in language model capabilities, offering robust performance across various domains and applications.