Model Releases
Ministral 3 3B: The Edge AI Revolution Arrives
Mistral AI's newest 3B parameter model brings vision capabilities to local edge devices under Apache 2.0. Perfect for phones and drones.
December 2, 2025
Model ReleaseMinistral 3 3B
Introduction
The landscape of artificial intelligence has shifted dramatically in 2025, moving away from massive cloud-dependent models toward efficient, local inference. Mistral AI has capitalized on this trend with the release of Ministral 3 3B on December 2nd, 2025. This new language model is designed specifically for edge computing, enabling developers to deploy sophisticated AI directly onto consumer hardware without relying on expensive cloud APIs. By combining a compact 3B parameter count with advanced vision capabilities, Mistral is democratizing AI for mobile and embedded systems.
This release marks a significant step forward in the industry's push toward artificial general intelligence, as noted by industry analysts tracking the most innovative companies in AI for 2025. Unlike previous iterations that required high-end GPUs, Ministral 3 3B is optimized to run smoothly on standard laptops, smartphones, and even drones. This shift allows for real-time processing of visual data and code generation directly on the device, reducing latency and enhancing user privacy.
- Released: December 2, 2025
- Parameters: 3 Billion
- License: Apache 2.0
- Capabilities: Text + Vision
- Target: Edge Devices
Key Features & Architecture
Ministral 3 3B is engineered with efficiency at its core. The architecture utilizes a dense transformer structure optimized for low-latency inference, ensuring that complex reasoning tasks do not bog down mobile processors. It features a native multimodal input layer that allows the model to process images alongside text prompts seamlessly. This vision capability is crucial for applications like drone navigation or mobile document analysis, where the model must understand visual context instantly.
The model supports a context window of 128k tokens, allowing it to handle extensive documents or video transcripts without losing coherence. Furthermore, the Apache 2.0 license ensures that developers can modify, distribute, and commercialize the model without restrictive clauses. This open-source approach fosters a vibrant ecosystem where community improvements can be rapidly integrated into the core model weights.
- Context Window: 128k tokens
- Input Types: Text + Vision
- License: Apache 2.0
- Quantization Support: INT4, FP4
- Inference Speed: 40 tokens/sec on CPU
Performance & Benchmarks
Despite its small footprint, Ministral 3 3B delivers competitive performance metrics compared to larger models. In the MMLU benchmark, it scores 68.5%, which is a 5% improvement over the previous 3B parameter baseline. On HumanEval, the coding benchmark, it achieves 62% accuracy, demonstrating strong capabilities for local development assistance. These scores are particularly impressive given the model's ability to run entirely on local hardware without external assistance.