ministral-3b-2512

Overview. ministral-3b-2512 is Mistral AI's Ministral 3 3B Instruct checkpoint from the December 2025 release: a dense 3.4B language model paired with a 0.4B Pixtral-type vision encoder, for a combined 3.85B parameters. It accepts text and images and returns text, with a 262,144-token context window, native function calling, and public weights under Apache-2.0. Developers reach for it when a small multimodal instruct model is enough and local or customer-managed weight deployment matters.

Key strengths. The useful signal here is architectural rather than leaderboard-based: no MMLU, Arena, or HumanEval figures are published for this checkpoint in the research used for this page. What is confirmed is a long 256k context window on a 3B-class dense model, native vision via the Pixtral-type encoder, native function calling, and an instruct post-train that accepts a system prompt. Apache-2.0 open weights make the same checkpoint available for self-hosting as well as hosted Inference.

When to use it. Use this for simple, latency-sensitive work — classification, short drafting, image captioning, and light tool calls — when a realtime 3B-class model is the right fit and multi-hop reasoning is not required. Prefer a larger Ministral sibling or a dedicated reasoning checkpoint when accuracy on constrained tasks matters more than size.

Limitations. This is an instruct checkpoint, not a reasoning model: it does not emit extended thinking traces, and Mistral AI ships chain-of-thought work on a separate Ministral 3 3B Reasoning release. The vendor-stated knowledge cutoff is 1 October 2023. There is no documented audio input or output and no native web search. JSON-schema constrained decoding is not documented for this checkpoint, while the shipped chat template renders multiple tool calls in a single turn.

Sovereignty note. Mistral AI is a French creator headquartered in Paris; it is not a US-parented company, though a US affiliate means CLOUD Act exposure is not categorically excluded. This release is open weights under Apache-2.0, so teams can run it in their own environment. On Stav it is served through the Mistral endpoint; confirm the endpoint's data-residency region and serving tier for regulated workloads. The vendor system prompt does not state a training-on-input policy for hosted calls.