ministral-14b-2512

Overview. Ministral 3 14B Instruct (2512) is Mistral AI's dense multimodal model, pairing a 13.5B language model with a 0.4B Pixtral-type vision encoder so it accepts text and images and returns text. It was released on 2025-12-02 under Apache-2.0 with public weights, as the instruction-tuned member of the Ministral 3 family. Developers reach for it when they want a compact model they can self-host, with a 262,144-token context window and native tool calling.

Key strengths. The vendor benchmarks on the model card put MATH Maj@1 at 0.904 and WildBench at 68.5, with Arena Hard at 0.551 — math and general chat quality that reads strong for a dense 14B model. The same card reports GPQA Diamond 0.57 and MMMU-Pro 0.50, suggesting the reasoning and vision paths hold up beyond text-only evaluation. Native function calling with tools and tool_choice is supported, along with structured outputs driven by JSON schema, and the tokenizer covers English, French, Spanish, German, Italian, Portuguese, Dutch, Chinese, Japanese, Korean and Arabic.

When to use it. Use this model for interactive work where roughly one-second first-token latency is acceptable: document and image question answering, structured extraction into JSON, and single-step tool calls. The 262,144-token context window suits long-document tasks that would otherwise need chunking. Keep it away from multi-step agentic loops — its agentic scores sit well below its general and math results, so orchestration-heavy workflows belong on a stronger reasoning model.

Limitations. The knowledge cutoff shipped in the vendor's own system prompt is 2023-10-01, so anything more recent may fall outside its training data. This is the Instruct checkpoint, not a reasoning model: it exposes no reasoning trace, and Mistral distributes a separate reasoning checkpoint for that role. Agentic reliability in multi-step tool use is the weakest measured area relative to its general capability.

Sovereignty note. Mistral AI is a French creator, and the Ministral 3 14B Instruct weights are published under Apache-2.0, so the model can be self-hosted or run on customer-managed infrastructure with no US CLOUD Act exposure. The active Stav endpoints for this model are operated from France and Switzerland.