Mistral AI Launches Mistral Large 2: Advanced AI Model

Today, Mistral AI announced the launch of Mistral Large 2, the new generation of its flagship model. Mistral Large 2 marks a significant upgrade from its predecessor, boasting enhanced capabilities in code generation, mathematics, and reasoning. This model also offers stronger multilingual support and advanced function-calling capabilities, setting a new standard in AI performance.

Enhanced Capabilities and Features

Cost Efficiency and Performance
Mistral Large 2 continues to push the boundaries of cost efficiency, speed, and overall performance. It is accessible via la Plateforme, enriched with features that facilitate the development of innovative AI applications.

Multilingual and Coding Proficiency
With a 128k context window, Mistral Large 2 supports dozens of languages, including French, German, Spanish, Italian, Portuguese, Arabic, Hindi, Russian, Chinese, Japanese, and Korean. It also covers over 80 coding languages such as Python, Java, C, C++, JavaScript, and Bash.

Single-Node Inference and Licensing
Designed with single-node inference for long-context applications in mind, Mistral Large 2’s 123 billion parameters allow it to run at large throughput on a single node. Released under the Mistral Research License, it is available for research and non-commercial use, with a commercial license required for self-deployment.

Performance Metrics

Evaluation Metrics and Benchmarks
Mistral Large 2 sets new benchmarks for performance and cost-efficiency. On MMLU, the pretrained version achieves an accuracy of 84.0%, positioning itself favorably against open models in terms of performance and cost.

Code-Centric Training
Building on the success of Codestral 22B and Codestral Mamba, Mistral Large 2 was trained on a substantial amount of code, outperforming its predecessor and matching the performance of leading models such as GPT-4o, Claude 3 Opus, and Llama 3 405B.

Reasoning and Accuracy
A significant effort was made to enhance Mistral Large 2’s reasoning capabilities. The model has been fine-tuned to reduce the generation of factually incorrect or irrelevant information, improving its reliability and accuracy. It is also trained to acknowledge when it lacks sufficient information to provide a confident answer.

Instruction-Following and Conversational Capabilities

Enhanced Communication Skills
The instruction-following and conversational abilities of Mistral Large 2 have seen significant improvements. The model excels in following precise instructions and handling long multi-turn conversations. Benchmarks such as MT-Bench, Wild Bench, and Arena Hard reflect its superior performance.

Conciseness in Business Applications
Recognizing the importance of conciseness in business applications, Mistral Large 2 has been optimized to generate succinct and relevant responses. This enhancement facilitates quicker interactions and cost-effective inference.

Multilingual Performance

Multilingual Benchmarks
A large fraction of business use cases involve working with multilingual documents. Mistral Large 2 was trained on extensive multilingual data and excels in multiple languages, including English, French, German, Spanish, Italian, Portuguese, Dutch, Russian, Chinese, Japanese, Korean, Arabic, and Hindi. The model's performance on the multilingual MMLU benchmark demonstrates its proficiency.

Function Calling and Business Applications

Enhanced Function Calling Skills
Mistral Large 2 is equipped with advanced function calling and retrieval skills, enabling it to handle both parallel and sequential function calls effectively. This makes it a powerful tool for complex business applications.

Availability and Platform Integration

La Plateforme Access
Mistral Large 2 is available on la Plateforme under the name mistral-large-2407, with versioning aligned to the YY.MM system. Users can test it on le Chat, and weights for the instruct model are hosted on HuggingFace.

Consolidation and Fine-Tuning Capabilities
The offering on la Plateforme is being consolidated around two general-purpose models, Mistral Nemo and Mistral Large, and two specialist models, Codestral and Embed. Fine-tuning capabilities have been extended to Mistral Large, Mistral Nemo, and Codestral.

Strategic Partnerships and Expansion

Global Cloud Partnerships
Mistral AI is expanding its partnership with Google Cloud Platform, making its models available on Vertex AI via a Managed API. Mistral AI’s top models are now accessible on Vertex AI, Azure AI Studio, Amazon Bedrock, and IBM watsonx.ai.

New Competitive Landscape
With the release of Mistral Large 2, Mistral AI introduces a model that matches or exceeds the performance of recent offerings from OpenAI and Meta, despite having fewer parameters. The model’s efficiency and impressive benchmarks place additional pressure on leading closed-AI companies like OpenAI, Anthropic, and Google.