ministral-8b-2512

Overview. Ministral 3 8B Instruct 2512 is a dense, instruction-tuned model from Mistral AI, released in December 2025 as part of the Ministral 3 family. It accepts text and image input, returns text, and ships with a 262,144-token context window. Developers can reach for it when they want a compact multimodal model whose weights are published openly.

Key strengths. The model pairs a language model of about 8.4B parameters with a 0.4B vision encoder, so image analysis comes with the model and needs no separate pipeline. Its 262,144-token context window is large for a model of this size. The Apache-2.0 weights are public and ungated. The Stav catalogue holds no benchmark scores for this checkpoint, so we make no benchmark claims here.

When to use it. Use it for chat and assistant workloads that mix text and images, such as describing screenshots or extracting details from photos. It also suits long-context prompts where a larger model is more than the task needs. The Smart Router can treat it as a fit for simple to medium tasks. Choose it when you want permissively licensed weights you can download and run yourself.

Limitations. Output is text only, and there is no audio input or output. Mistral publishes a separate Reasoning checkpoint, so this Instruct model is not that variant, and Base and Reasoning results should not be read as this model's results. The vendor's shipped system prompt gives a knowledge update date of October 2023, so recent events may be missing.

Sovereignty note. Mistral AI is the creator, and the weights are released under Apache-2.0, so you can run Inference on infrastructure you control. This model's training-on-input policy and endpoint data residency have not been confirmed, so check the endpoint you select.