README.md

# Ministral-3-3B-Turbo

Run Ministral-3-3B on Qualcomm NPU with NexaSDK and turbo optimization.

## Quickstart

Install nexaSDK and create a free account at [sdk.nexa.ai](https://sdk.nexa.ai)

Activate your device with your access token:

```bash
nexa config set license '<access_token>'
```

Run the model locally in one line:

```bash
nexa infer NexaAI/Ministral-3-3B-turbo-npu
```

## Model Description
**Ministral-3-3B-Instruct-2512** is the instruction-tuned variant of Mistral AI’s smallest Ministral 3 model: a compact multimodal language model combining a ~3.4B-parameter language core with a 0.4B-parameter vision encoder.  
It is post-trained in FP8 for instruction-following, making it well-suited for chat-style agents, tool use, and grounded reasoning on both text and images.  
With a large 256k context window and efficient edge-oriented design, it targets real-time use on GPUs and other resource-constrained hardware.

## Features
- **Multimodal (vision + text)**: Understands and reasons over images alongside text in a single conversation.
- **Instruction-tuned**: Optimized for following natural-language instructions, chat, and assistant-style workflows.
- **Agentic capabilities**: Native support for function calling and structured JSON-style outputs for tool and API orchestration.
- **Large context window**: Up to **256k tokens** for long documents, multi-step workflows, and complex sessions.
- **Edge-optimized FP8 weights**: FP8 checkpoint designed for efficient deployment and serving, including on a single modern GPU.
- **Multilingual**: Supports dozens of languages, including English, French, Spanish, German, Italian, Portuguese, Dutch, Chinese, Japanese, Korean, and Arabic.
- **Part of the Ministral 3 family**: Seamlessly aligned with 3B/8B/14B base, instruct, and reasoning variants for scalable deployments.

## Use Cases
- **Vision + language assistants**
  - Image captioning and explanation (UI screenshots, photos, diagrams)
  - Multimodal Q&A (e.g., “describe this chart and summarize its implications”)
- **Lightweight agents and tools**
  - Function-calling workflows (retrieval, calculators, external APIs)
  - JSON-structured responses for downstream automation
- **Text understanding & generation**
  - Classification, tagging, routing, and extraction from long documents
  - Short-form copywriting, drafting, and rewriting across multiple languages
- **Edge & low-resource deployments**
  - On-device or near-edge assistants where latency, context length, and cost matter
  - Local/private workloads that benefit from a small yet capable multimodal model

## Inputs and Outputs

**Inputs**
- **Text-only prompts**
  - Single-turn or multi-turn chat-style conversations (`system`, `user`, `assistant` roles).
  - Long-context inputs up to the model’s context limit (e.g., documents, logs, transcripts).
- **Multimodal prompts**
  - One or more images (e.g., URLs or image tensors) combined with text.
- **Structured tool schemas**
  - Function / tool definitions for agentic workflows (JSON schemas describing functions and parameters).

**Outputs**
- **Generated text**
  - Answers, explanations, step-by-step reasoning, summaries, and creative content.
- **Multimodal-aware responses**
  - Text grounded in the provided images (descriptions, comparisons, localized details).
- **Structured tool calls**
  - JSON-like tool call objects for function execution and programmatic integration.
- **Logits / probabilities (advanced)**
  - For users accessing the raw model via low-level APIs, token-level scores for custom decoding or research.

## License
This repo is licensed under the Creative Commons Attribution–NonCommercial 4.0 (CC BY-NC 4.0) license, which allows use, sharing, and modification only for non-commercial purposes with proper attribution. All NPU-related models, runtimes, and code in this project are protected under this non-commercial license and cannot be used in any commercial or revenue-generating applications. Commercial licensing or enterprise usage requires a separate agreement. For inquiries, please contact `[email protected]`