Mistral AI has unveiled Mistral Medium 3, an affordable model on the smaller side, designed to deliver performance comparable to large and often costly foundation models. According to the company, Mistral Medium 3 establishes a new model category that balances cutting-edge performance, significantly lower operational costs, and simple deployment, enabling enterprise users to integrate the model easily into their existing systems.

The new model delivers at least 90% of Claude Sonnet 3.7's performance across several popular benchmarks testing everything from coding and math to instruction following and multimodal capabilities, while being 8x less expensive ($0.4 input/$2 output per million tokens) than that same model. According to company evaluations, Medium 3 outperforms leading openly available rivals like Llama 4 Maverick, and Cohere Command A, while beating contenders such as DeepSeek v3 on pricing for API and self-deployments.

To strengthen its case for Medium 3's performance, Mistral AI provides results for third-party human preference evaluations based on the claim that they better reflect performance in real-world scenarios. Expectedly, in evaluations for coding tasks, the scores reveal a strong preference (60%) for Claude Sonnet 3.7. However, it is also surprising that, while the evaluations also show over 60% preference for DeepSeek 3.1, Medium 3 is exactly tied with GPT-4o and is very strongly preferred over Command-A (69.23%) and Llama 4 Maverick (with a staggering 81.82% preference rate). Human preference evaluation results across several domains, including widely spoken languages, multimodal capabilities, and coding, show an overall user preference for Medium 3 over Llama 4 Maverick.

Credit: Mistral AI
Credit: Mistral AI

Mistral Medium 3 can be deployed on any cloud environment, including self-hosting solutions with at least four GPUs. When combined with Mistral's applied AI solutions, this flexibility enables comprehensive enterprise integration, including continuous pretraining, full fine-tuning, and knowledge base blending.

The model is available immediately via Mistral La Plateforme and Amazon SageMaker. Support on IBM WatsonX, NVIDIA NIM, Azure AI Foundry, and Google Cloud Vertex is upcoming.

Mistral AI also teased an upcoming "large" model release in the coming weeks, hinting that it will be even more impressive than Medium 3, which already outperforms larger and costlier systems like Llama 4 Maverick.