mistral-small-2603

Overview. Mistral Small 4 is an open-weights Mixture-of-Experts model from Mistral AI, released on 2026-03-16 under Apache-2.0. It carries 119B total parameters with 6.5B activated per token across 128 experts, of which 4 are active per token, and it accepts text and image input against a 262,144-token context window. It is built as a hybrid checkpoint: by default it answers in instruct mode, and callers can opt into a visible reasoning trace by raising the reasoning effort. Developers reach for it when they want instruction following, vision, long-context handling and optional reasoning in one deployable checkpoint rather than several specialist models.

Key strengths. The model's distinguishing property is the hybrid design: reasoning effort of 'none' returns a direct instruct-style answer, while 'high' produces a full thinking chunk, so the same checkpoint covers both fast and deliberate workloads. Input modality is text plus image, with text-only output, and the architecture is a Mistral3-family conditional-generation stack. Published third-party measurement is thin, but Artificial Analysis reports a quality index of 9 in the model's default mode (rising to 11.3 with reasoning enabled) and a coding index of 26.6, the latter recorded with reasoning enabled rather than in default mode. Context handling extends to 262,144 tokens, which suits long documents and large in-context codebases.

When to use it. Use Mistral Small 4 for medium-complexity chat, writing and code tasks where an interactive latency class is acceptable and the workload does not require frontier-level reasoning. It fits workflows that mix text and images, such as screenshot or diagram understanding, and long-context tasks that would otherwise need chunking inside the 262,144-token window. The opt-in 'high' reasoning effort is the right setting when a task benefits from an intermediate thinking step, while the default instruct mode is the faster path for routine requests. It is a reasonable fit where deployment flexibility matters, since the weights are openly licensed and can be served on infrastructure the operator controls.

Limitations. Reasoning effort supports only two documented levels, 'none' and 'high', so there is no intermediate rung to tune deliberation against latency. Input is limited to text and image, with no audio and no audio output. Mistral AI does not publish a training-data cutoff for this model, so recency of world knowledge cannot be stated precisely, and the available benchmark reporting is limited to third-party quality and coding indices rather than the wider set of published evaluations.

Sovereignty note. Mistral Small 4 is built by Mistral AI, a French company, and its weights are open under Apache-2.0, so it can be run on infrastructure of your choice. Where requests run, and whether inputs are retained or used for training, depends on the endpoint serving it; see the endpoint facts on this page.