Mistral Large 2: The New Benchmark in AI Performance and Multilingual Proficiency

The latest generation of AI models continues to redefine the landscape of cost efficiency, speed, and performance. At the forefront of this evolution is Mistral Large 2, now available on la Plateforme. This model is enriched with new features designed to facilitate the development of innovative AI applications, making it a significant player in the AI space.

Unmatched Performance and Cost Efficiency

Mistral Large 2, boasting an impressive 128k context window, supports a broad range of languages including French, German, Spanish, Italian, Portuguese, Arabic, Hindi, Russian, Chinese, Japanese, and Korean. Additionally, it supports over 80 coding languages such as Python, Java, C, C++, JavaScript, and Bash. Designed for single-node inference with long-context applications, Mistral Large 2’s 123 billion parameters enable high throughput on a single node.

This model is released under the Mistral Research License, allowing for usage and modification for research and non-commercial purposes. For commercial usage requiring self-deployment, a Mistral Commercial License is necessary.

In terms of performance, Mistral Large 2 sets a new standard, particularly on the MMLU benchmark where it achieves an accuracy of 84.0%, positioning itself favorably on the performance/cost Pareto front of open models.

Superior Code and Reasoning Capabilities

Building on the experience from Codestral 22B and Codestral Mamba, Mistral Large 2 has been trained on a large proportion of code. It significantly outperforms its predecessor, Mistral Large, and rivals leading models like GPT-4, Claude 3 Opus, and Llama 3 405B.

A key focus during training was to enhance the model’s reasoning capabilities and minimize the tendency to generate plausible-sounding but incorrect information. This fine-tuning ensures that Mistral Large 2 provides reliable and accurate outputs, acknowledging when it cannot find solutions or lacks sufficient information to provide a confident answer. This commitment to accuracy is evident in its improved performance on popular mathematical benchmarks.

Enhanced Instruction Following and Alignment

Mistral Large 2 excels in instruction-following and conversational capabilities, handling precise instructions and long multi-turn conversations with ease.

Its performance on benchmarks like MT-Bench, Wild Bench, and Arena Hard highlights these capabilities. Importantly, the model is designed to generate concise and to-the-point responses, essential for business applications where quick interactions and cost-effective inference are paramount.

Multilingual Proficiency

In today’s globalized business environment, working with multilingual documents is crucial. Unlike many English-centric models, Mistral Large 2 has been trained on a large proportion of multilingual data. It performs exceptionally well in languages such as English, French, German, Spanish, Italian, Portuguese, Dutch, Russian, Chinese, Japanese, Korean, Arabic, and Hindi. Its performance on the multilingual MMLU benchmark surpasses that of the previous Mistral Large, Llama 3.1 models, and Cohere’s Command R+.

Advanced Tool Use and Function Calling

Mistral Large 2 is equipped with enhanced function calling and retrieval skills, trained to proficiently execute both parallel and sequential function calls. This capability makes it an ideal engine for complex business applications.

Access and Integration

Mistral Large 2 is available today via la Plateforme under the name mistral-large-2407, and can be tested on le Chat. The model is available under version 24.07 and the API name mistral-large-2407. The instruct model weights are also hosted on HuggingFace.

La Plateforme consolidates its offerings around two general-purpose models, Mistral Nemo and Mistral Large, and two specialist models, Codestral and Embed. Older models remain available for deployment and fine-tuning using the Mistral SDK.

Cloud Integration

Mistral AI is expanding its partnerships with leading cloud service providers, making Mistral Large 2 accessible to a global audience. This includes a collaboration with Google Cloud Platform to offer Mistral AI’s models on Vertex AI via a Managed API. Additionally, Mistral AI’s top models are available on Azure AI Studio, Amazon Bedrock, and IBM watsonx.ai.

Conclusion

Mistral Large 2 represents a significant advancement in AI model performance and cost efficiency. Its superior language support, coding proficiency, reasoning capabilities, and enhanced instruction-following skills make it a formidable competitor to top models like GPT-4, Claude 3 Opus, and Llama 3.1.

With its broad accessibility through leading cloud platforms and dedicated support on la Plateforme, Mistral Large 2 is set to play a crucial role in the development of next-generation AI applications.

If you want more updates related to AI, subscribe to our Newsletter