Mistral Large 2 (also known as Mistral-Large-Instruct-2407) represents a significant advancement in large language model technology. With 123 billion parameters, this dense LLM is designed for advanced reasoning, knowledge processing, and coding capabilities while being optimized for single-node inference with long-context applications. The model features an expansive 128k context window, enabling it to process and maintain coherence across lengthy documents and conversations.
One of the model's standout features is its extensive multilingual capabilities, supporting dozens of languages including English, French, German, Spanish, Italian, Chinese, Japanese, Korean, Portuguese, Dutch, Polish, Arabic, Hindi, and Russian. This broad language support is complemented by proficiency in over 80 programming languages, ranging from commonly used languages like Python and Java to more specialized ones such as Swift and Fortran.
Performance and Capabilities
Mistral Large 2 demonstrates impressive performance across a wide range of benchmarks. On the MMLU benchmark, it achieves an 84% overall score, with particularly strong performance in multilingual subsets ranging from 60.1% (Korean) to 82.8% (French). The model shows exceptional capabilities in code generation and reasoning tasks, achieving 92% on HumanEval, 87% on HumanEval Plus, 80% on MBPP Base, and 69% on MBPP Plus.
Mathematical reasoning is another area where the model excels, scoring 93% on GSM8K, 70% on Math Instruct (0-shot, no CoT), and 71.5% on Math Instruct (0-shot, CoT). In instruction following and conversational capabilities, the model achieves scores of 8.63 on MT Bench, 56.3 on Wild Bench, and 73.2 on Arena Hard.
A key focus during the model's development was minimizing hallucinations through fine-tuning for more cautious and discerning responses. The model has been trained to acknowledge when it lacks sufficient information, contributing to its reliability in real-world applications.
Technical Implementation and Usage
The model can be utilized through two primary frameworks: mistral_inference (recommended) and transformers. Implementation requires substantial computational resources, with over 300GB of VRAM needed for operation. The model features best-in-class agentic capabilities with native function calling and JSON output.
Mistral Large 2 is released under the Mistral Research License, which permits usage and modification for research and non-commercial purposes. The model is available on various platforms and can be downloaded via the Hugging Face Hub, though commercial use requires a separate license.
References
- Mistral AI Release Blog Post - Official announcement and detailed model information
- Mistral Inference GitHub - Technical implementation guide
- Hugging Face Transformers - Alternative implementation framework
- Function Calling Guide - Detailed technical documentation
- Mistral Research License - Legal usage terms and conditions
- Privacy Policy - Data handling and privacy information
