Mistral Small 4 belongs to the mistral-small family and is distributed as an open-weights 119B mixture-of-experts model, a lineage that third-party deployment reports have begun to explore on high-memory hardware. A community post on the NVIDIA developer forums documents the model running as a 119B MoE setup with the SGLang inference framework on DGX Spark, providing early practical evidence that the weights are accessible for self-hosted experimentation rather than locked behind closed APIs. The model is also listed on the NVIDIA NGC catalog under the mistralai organization's NIM containers, giving operators an established channel for pulling the checkpoint into containerized environments alongside other NIM-supported releases.
Within the mistral-small family, Mistral Small 4 sits in a mid-size open-weights tier aimed at teams that want reasoning, vision input, and tool-calling behavior without committing to a larger frontier deployment. The 119B MoE design lets active parameters stay modest per token while still drawing on a broad expert pool, which suits agents and retrieval-heavy applications where structured outputs and temperature control matter as much as raw throughput. For practitioners, the practical fit is a model that can be hosted locally or through a third-party cloud, integrated with attachments and function-calling pipelines, and upgraded over time as the mistral-small line continues to evolve beyond the 2603 release point.
