On May 7, 2025, Mistral AI, a leading French artificial intelligence startup, officially unveiled its latest language model, Mistral Medium 3. This new model sets a new benchmark in the AI landscape by delivering state-of-the-art performance at a fraction of the cost compared to existing competitors. With its advanced multimodal capabilities, extended context length, and enterprise-ready features, Mistral Medium 3 is positioned to become a game-changer for businesses seeking powerful yet cost-effective AI solutions.
Cutting-Edge Performance with Cost Efficiency
Mistral Medium 3 introduces a new class of models that balances top-tier performance with significant cost reductions. Benchmarks show that it achieves or surpasses 90% of the performance of Anthropic’s Claude Sonnet 3.7 across a variety of tasks, including coding, STEM reasoning, and multimodal understanding. Remarkably, it delivers this level of performance at approximately one-eighth the operational cost, priced at $0.40 per million input tokens and $2 per million output tokens.
In addition to outperforming Claude Sonnet 3.7, Mistral Medium 3 also surpasses major open models such as Meta’s Llama 4 Maverick and enterprise models like Cohere Command A. This combination of high accuracy and low cost makes it an attractive choice for enterprises aiming to scale AI applications without prohibitive expenses.
Versatile Multimodal and Long-Context Capabilities
One of the standout features of Mistral Medium 3 is its multimodal capability, enabling it to process and understand both text and visual inputs. This expands its applicability across diverse domains, from document analysis to image understanding.
Moreover, the model supports an extended context window of up to 128,000 tokens, allowing it to handle long documents and complex conversations with ease. This long-context capability is critical for enterprise use cases such as legal document review, technical manuals, and multi-turn dialogue systems.
Enterprise-Grade Deployment and Flexibility
Mistral Medium 3 is designed with enterprise needs in mind, providing flexible deployment options including hybrid, on-premises, and virtual private cloud (VPC) environments. The model can efficiently run on as few as four NVIDIA H100 GPUs, reducing infrastructure costs and making it accessible for organizations with limited hardware resources.
Its compatibility with major cloud platforms such as Amazon SageMaker, IBM WatsonX, Microsoft Azure AI Foundry, NVIDIA NIM, and Google Cloud Vertex AI ensures seamless integration into existing enterprise workflows.
Strong Benchmark Results Across Domains
The model has demonstrated exceptional results in various benchmark tests:
-
Coding and STEM Tasks: Achieved a HumanEval 0-shot score of 0.921, matching Claude Sonnet 3.7 and outperforming Llama 4 Maverick.
-
Mathematical Reasoning: Scored 0.91 on Math500 Instruct 0-shot, surpassing GPT-4o and other competitors.
-
Multimodal Understanding: Outperformed GPT-4o and Llama 4 Maverick in the MMMU 0-shot benchmark.
-
Long-Context Processing: Matched or exceeded GPT-4o in RULER 32K and 128K benchmarks, demonstrating strong comprehension over extended inputs.
These results underscore Mistral Medium 3’s ability to handle complex, real-world tasks with high accuracy.
Broad Language Support and Customization
Mistral Medium 3 supports over 40 languages, including English, Japanese, Hindi, Swedish, Chinese, Catalan, and Greek, making it suitable for global applications. Enterprises can also customize the model through post-training fine-tuning to better fit specific business needs, enhancing performance in specialized domains.
Conclusion
Mistral Medium 3 represents a significant advancement in enterprise AI, combining frontier-class performance with unprecedented cost efficiency and deployment flexibility. Its multimodal capabilities and long-context understanding open new possibilities for industries such as finance, healthcare, energy, and more. By offering a powerful yet affordable solution, Mistral AI is poised to accelerate the adoption of AI across enterprises worldwide, enabling smarter, faster, and more scalable AI-driven innovation.