Mistral AI SAS, a French artificial intelligence company, has released Mistral Small 4, a new iteration in its model series that integrates reasoning, multimodal, and agentic coding features into a single architecture. This release combines the capabilities of the previous Magistral, Pixtral, and Devstral models, allowing users to perform various tasks within one system. The model utilizes a Mixture of Experts architecture with 119 billion total parameters and is available under the Apache 2.0 open-source license. Mistral has also joined the NVIDIA Nemotron Coalition as a founding member.
Technical specifications for Mistral Small 4 include a 256k context window and support for both text and image inputs. A new “reasoning on demand” parameter allows for adjustments between standard response speeds and more intensive, step-by-step processing. In terms of infrastructure, the model can be deployed on hardware starting with a single NVIDIA DGX B200, and it shows a 40% reduction in end-to-end completion time compared to the previous version.
In their statement, they share that internal benchmark data indicates that Mistral Small 4 performs similarly to or better than GPT-OSS 120B across several metrics while generating shorter outputs. According to the developer, this efficiency reduces resource consumption for tasks such as codebase exploration and document analysis. The model is currently accessible through the Mistral API, Hugging Face, and as an NVIDIA NIM.
