Mistral Medium 3 Outperforms Llama 4; Open Model Launch Planned Next

Mistral AI unveiled Mistral Medium 3, a new AI model that claims to balance cutting-edge performance with cost-effectiveness while outperforming competitors like Meta's Llama 4 Maverick in benchmark tests.

The company stated in the announcement that the new model is specifically designed for enterprise deployment and excels in coding, STEM, and multimodal tasks. According to the company, Mistral Medium 3 achieves over 90% of Claude Sonnet 3.7's benchmark scores at significantly lower pricing—$0.40 per million tokens for input and $2 for output.

This comes right after its open-source model, Mistral Small 3.1’s release. The new model builds on Mistral Small 3, with improved text performance, multimodal understanding, and an expanded context window of up to 128k tokens. Mistral claimed that the model outperforms comparable models like Gemma 3 and GPT-4o mini while delivering inference speeds of 150 tokens per second. 

Mistral Medium 3 can be deployed in hybrid or on-premise environments with support for continuous pretraining and enterprise system integration. The company reports that early adopters in finance, energy, and healthcare sectors are already using it for personalised customer service and complex data analysis.

Despite its medium size, the model reportedly outperforms several larger competitors in both API and self-hosted formats. It can run on systems with as few as four GPUs, making deployment more accessible for organisations with varying infrastructure capabilities.

In third-party human evaluations focused on real-world scenarios, Mistral Medium 3 particularly shines in coding tasks, surpassing some significantly larger models.

The company claims that on benchmarks, Mistral Medium 3 outperforms Cohere Command A alongside Llama 4 Maverick, while beating DeepSeek v3 on pricing in both API and self-deployed scenarios.

The model is now available on Mistral's own platform and Amazon SageMaker, with upcoming support on Azure AI, Google Cloud, IBM WatsonX, and NVIDIA NIM.

Interestingly, for future releases, Mistral confirmed that a larger open model is in development.