Ministral 3B
Version: 1
Ministral 3B is a state-of-the-art Small Language Model (SLM) optimized for edge computing and on-device applications. As it is designed for low-latency and compute-efficient inference, it it also the perfect model for standard GenAI applications that have real-time requirements and high-volume.
Number of Parameters: 3,6 billions
Ministral 3B and Ministral 8B set a new frontier in knowledge, commonsense, reasoning, function-calling, and efficiency in the sub-10B category, and can be used or tuned to a variety of uses, from orchestrating agentic workflows to creating specialist task workers. Both models support up to 128k context length (currently 32k on vLLM) and Ministral 8B has a special interleaved sliding-window attention pattern for faster and memory-efficient inference.
Use cases
Our most innovative customers and partners have increasingly been asking for local, privacy-first inference for critical applications such as on-device translation, internet-less smart assistants, local analytics, and autonomous robotics. Les Ministraux were built to provide a compute-efficient and low-latency solution for these scenarios. From independent hobbyists to global manufacturing teams, les Ministraux deliver for a wide variety of use cases. Used in conjunction with larger language models such as Mistral Large, les Ministraux are also efficient intermediaries for function-calling in multi-step agentic workflows. They can be tuned to handle input parsing, task routing, and calling APIs based on user intent across multiple contexts at extremely low latency and cost. Source: Un Ministral, des Ministraux - Introducing the world’s best edge models.Content Filtering
Prompts and completions are passed through a default configuration of Azure AI Content Safety classification models to detect and prevent the output of harmful content. Learn more about Azure AI Content Safety . Configuration options for content filtering vary when you deploy a model for production in Azure AI; learn more .Source: Un Ministral, des Ministraux - Introducing the world’s best edge models.
We demonstrate the performance of les Ministraux across multiple tasks where they consistently outperform their peers. We re-evaluated all models with our internal framework for fair comparison.
Pretrained Models
Knowledge & Commonsense
| Model | MMLU | AGIEval | Winogrande | Arc-c | TriviaQA |
|---|---|---|---|---|---|
| Gemma 2 2B | 52.4 | 33.8 | 68.7 | 42.6 | 47.8 |
| Llama 3.2 3B | 56.2 | 37.4 | 59.6 | 43.1 | 50.7 |
| Ministral 3B | 60.9 | 42.1 | 72.7 | 64.2 | 56.7 |
| Mistral 7B | 62.4 | 42.5 | 74.2 | 67.9 | 62.5 |
| Llama 3.1 8B | 64.7 | 44.4 | 74.6 | 46.0 | 60.2 |
| Ministral 8B | 65.0 | 48.3 | 75.3 | 71.9 | 65.5 |
Code and Math
| Model | HumanEval (pass@1) | GSM8K (maj@8) |
|---|---|---|
| Gemma 2 2B | 20.1 | 35.5 |
| Llama 3.2 3B | 29.9 | 37.2 |
| Ministral 3B | 34.2 | 50.9 |
| Mistral 7B | 26.8 | 51.3 |
| Llama 3.1 8B | 37.8 | 61.7 |
| Ministral 8B | 34.8 | 64.5 |
Multilingual
| Model | French MMLU | German MMLU | Spanish MMLU |
|---|---|---|---|
| Gemma 2 2B | 41.0 | 40.1 | 41.7 |
| Llama 3.2 3B | 42.3 | 42.2 | 43.1 |
| Ministral 3B | 49.1 | 48.3 | 49.5 |
| Mistral 7B | 50.6 | 49.6 | 51.4 |
| Llama 3.1 8B | 50.8 | 52.8 | 54.6 |
| Ministral 8B | 57.5 | 57.4 | 59.6 |
Instruct Models
Chat/Arena (gpt-4o judge)
| Model | MTBench | Arena Hard | Wild bench |
|---|---|---|---|
| Gemma 2 2B | 7.5 | 51.7 | 32.5 |
| Llama 3.2 3B | 7.2 | 46.0 | 27.2 |
| Ministral 3B | 8.1 | 64.3 | 36.3 |
| Mistral 7B | 6.7 | 44.3 | 33.1 |
| Llama 3.1 8B | 7.5 | 62.4 | 37.0 |
| Gemma 2 9B | 7.6 | 68.7 | 43.8 |
| Ministral 8B | 8.3 | 70.9 | 41.3 |
Code and Math
| Model | MBPP (pass@1) | HumanEval (pass@1) | Math (maj@1) |
|---|---|---|---|
| Gemma 2 2B | 54.5 | 42.7 | 22.8 |
| Llama 3.2 3B | 64.6 | 61.0 | 38.4 |
| Ministral 3B | 67.7 | 77.4 | 51.7 |
| Mistral 7B | 50.2 | 38.4 | 13.2 |
| Llama 3.1 8B | 69.7 | 67.1 | 49.3 |
| Gemma 2 9B | 68.5 | 67.7 | 47.4 |
| Ministral 8B | 70.0 | 76.8 | 54.5 |
Model Specifications
Context Length131072
Quality Index0.45
LicenseCustom
Last UpdatedOctober 2024
Input TypeText
Output TypeText
PublisherMistral AI
Languages5 Languages
Related Models
