Ministral-3-3B-Reasoning-2512

Ministral-3-3B-Reasoning-2512 combines a language model with 3.4 billion parameters and a visual encoder with 0.4 billion parameters. The model is the result of a three-stage cascaded distillation process: the parent model Mistral Small 3.1 (24B) is progressively pruned to 14B, then to 8B, and finally to 3B, with knowledge distillation from a larger “teacher” at each stage. This approach preserves high generation quality while radically reducing the number of parameters and training compute costs. Despite its compact size, the model retains a full 0.4B‑parameter visual encoder, enabling image understanding on par with textual information. This makes Ministral-3-3B one of the few models in the sub‑4B category with built‑in multimodality.

The model is a full‑fledged reasoning version, post‑trained to handle tasks that require logical reasoning — mathematics, programming, natural sciences. It supports dozens of languages, provides strict adherence to system prompts, and offers agentic capabilities with native tool calling and JSON output. The 256k token context window is preserved even in this most compact version of the family. On the AIME24 math benchmark, Ministral-3-3B scores 0.775, outperforming Qwen3-VL-4B-Thinking with 0.729. On AIME25, the model scores 0.721 vs. 0.697 for the competitor. On GPQA Diamond it achieves 0.534 (Qwen3-VL-4B-Thinking – 0.601); thus the results are exceptionally high for a 3B model.

Recommendations for working with images are similar to those for larger models: an aspect ratio close to 1:1 and cropping when necessary for optimal performance. It is also advised to use the curated system prompt available at https://huggingface.co/mistralai/Ministral-3-3B-Reasoning-2512/blob/main/SYSTEM_PROMPT.txt , supplementing it with custom instructions to define the environment and use case, including guidance on effective tool usage in agentic systems. In multi‑turn dialogues, it is important to keep reasoning traces in the context. The recommended sampling temperature is 0.7 for most environments, and the tool set should be clearly defined and limited to the minimum necessary to avoid overloading the model.

Ministral-3-3B is ideal for lightweight, real‑time applications on edge devices or hardware with minimal resources. Key use cases include: image captioning, text classification, efficient real‑time translation, data extraction from unstructured sources, short content generation, and fine‑tuning / specialization for narrow tasks.

Announce Date: 31.10.2025
Parameters: 5B
Context: 263K
Layers: 26
Attention Type: Full Attention
Developer: Mistral AI
Transformers Version: 5.0.0.dev0
License: Apache 2.0

Use our pre-built public endpoints for free to test inference and explore Ministral-3-3B-Reasoning-2512 capabilities. You can obtain an API access token on the token management page after registration and verification.

Rent your own physically dedicated instance with hourly or long-term monthly billing.

We recommend deploying private instances in the following scenarios:

• maximize endpoint performance,
• enable full context for long sequences,
• ensure top-tier security for data processing in an isolated, dedicated environment,
• use custom weights, such as fine-tuned models or LoRA adapters.

Contact our dedicated neural networks support team at support@immers.cloud or send your request to the sales department at sale@immers.cloud.