Ministral 3 3B

Ministral 3 3B belongs to Mistral AI's Ministral3 efficient small model series, designed to balance compact size with practical capability for on-device and edge inference. The Instruct-2512 variant carries the Mistral3ForConditionalGeneration architecture and operates as a 3-billion-parameter vision-language model, accepting image and text inputs to produce text outputs. This positions it well for multimodal tasks like visual question answering and document understanding while keeping the footprint modest enough to run on constrained hardware, including Qualcomm's Snapdragon X2 Elite targets via llama.cpp runtimes, and supported through the NVIDIA NeMo AutoModel ecosystem for supervised fine-tuning recipes.

As an open-weights release, the Ministral-3-3B-Instruct-2512 checkpoint is publicly distributed on Hugging Face under the mistralai organization, with third-party community builds (for example, GGUF quantizations for local inference) making it accessible to independent developers. The combination of an Apache-2.0 license, fine-tuning documentation for domain adaptation such as MedPix medical imagery, and quantization-friendly formats supports flexible deployment from phones and tablets through workstation GPUs, suiting developers who need a small, adaptable multimodal model they can customize and self-host.

Mistralministral-3b-2512ministral

Open provider page Provider docs

Quick Info

Provider: Mistral
Model key: ministral-3b-2512
Release date: Dec 2, 2025
Last updated: Dec 2, 2025
Input modalities
Output modalities
Capabilities

Cost

Input token cost: $0.10
Output token cost: $0.10

Limits

Output tokens: 262,144 tokens
Context window: 131,072 tokens

Transparent token rates

Compare Ministral 3 3B pricing

Rates are shown per one million tokens. Combined means one million input plus one million output tokens.

Browse this family

Latest news about Ministral 3 3B

Mistral

Official sourceAnnouncement

Mistral AI announced Mistral 3 on December 2, 2025, introducing a new generation of models including three dense variants — 14B, 8B, and 3B — alongside Mistral Large 3, a sparse mixture-of-experts model with 41B active and 675B total parameters. All models are released under the Apache 2.0 license. The Ministral models are positioned as offering the best performance-to-cost ratio in their category, targeting efficient deployment for developers and enterprises. All Mistral 3 models, from Large 3 down to Ministral 3, were trained on NVIDIA Hopper GPUs. Mistral partnered with NVIDIA, vLLM, and Red Hat to optimize accessibility, releasing an NVFP4 checkpoint built with llm-compressor that enables efficient inference on Blackwell NVL72 systems and on 8×A100 or 8×H100 nodes. The release also debuts Mistral's first MoE since the Mixtral series, marking a substantial pretraining step forward for the company.