Ministral 3 3B Reasoning 2512

Ministral 3 3B Reasoning 2512 is the smallest member of the Ministral 3 family and a reasoning-focused post-trained variant aimed at math, coding, and other STEM tasks. Its design pairs a 3.4B-parameter language model with a 0.4B vision encoder, giving it image-understanding capability alongside standard text input and text output. The family is positioned for edge deployment across a broad range of hardware, with the 3B variant able to run locally at BF16 within roughly 16GB of VRAM and at well under 8GB of RAM or VRAM when quantized, making it suitable for lightweight, real-time applications on resource-constrained devices.

In practice, the model combines that compact multimodal footprint with practical agentic features, including native function calling and JSON output, strong adherence to system prompts, and broad multilingual coverage across English, French, Spanish, German, Italian, Portuguese, Dutch, Chinese, Japanese, Korean, and Arabic. Community distributions such as the Q4_K_M GGUF build available on Ollama weigh in around 3.0GB, reinforcing the edge-friendly sizing, while the open weights and agentic tooling support make it a flexible choice for on-device assistants, image captioning, text classification, and other small-scale reasoning or translation workloads.