Mistral Small 4

Mistral Small 4 is positioned as a mid-to-large open-weight generation in the Mistral family, surfaced through Infomaniak's hosted catalog with the cataloged API limit context window for both input and output. The model key in the catalog (mistralai/Mistral-Small-4-119B-2603) and an independent community write-up describe it as a 119B Mixture-of-Experts design, suggesting a sparse-activation architecture intended to balance capacity against inference cost rather than acting as a dense flagship. This MoE framing matters in practice: developers can target higher-quality reasoning and tool-assisted workflows while paying only for the experts actually activated per token, which is consistent with the model's positioning between Mistral's smaller open models and its larger frontier-tier releases.

A community benchmark thread on the NVIDIA DGX Spark forum documents practitioners running the model locally via SGLang, signaling that the weights are genuinely usable for self-hosting rather than being API-only, and giving the model a natural home in agentic and tool-calling pipelines. The the cataloged API limit context budget makes it well suited to long-document analysis, codebases, and multi-turn agent loops where tool inputs and retrieved context can accumulate quickly. For teams already invested in the Mistral ecosystem, the practical fit is a capable general-purpose reasoning model that handles text and image inputs at a mid-tier price point while remaining open-weight enough to deploy on accelerated hardware.