Mistral Small 3
2025-01-30 · Publisher : Mistral AI · Domain : Artificial Intelligence
Mistral AI releases Mistral Small 3, a 24-billion-parameter model optimized for latency, under the Apache 2.0 license, presented as comparable to Llama 3.3 70B while being more than 3 times faster.
What changes
- Return to the Apache 2.0 license for the Small range, with pre-trained and instruction-tuned checkpoints.
- Architecture with fewer layers than its competitors to reduce the time per forward pass: 150 tokens/s claimed.
- Neither RL training nor synthetic data: designed as a base for derived reasoning models.
- Targets: fast conversational assistants, low-latency function calling, domain fine-tuning, local inference.
### Specifications - 24 billion parameters; 32k-token context (documentation and Hugging Face model card). - Text only; languages: English, French, German, Spanish, Italian, Chinese, Japanese, Korean, Portuguese, Dutch, Polish, among others. - License: Apache 2.0. - Availability: La Plateforme, Hugging Face, Ollama, Kaggle, Together AI, Fireworks AI, IBM watsonx; NVIDIA NIM, SageMaker, Groq, Databricks and Snowflake announced. - API identifiers: `mistral-small-latest`, `mistral-small-2501` (retired on 11/30/2025). ### Pricing | Model | Input (USD / M tokens) | Output (USD / M tokens) | |---|---|---| | Mistral Small 25.01 | 0.1 | 0.3 | Pricing recorded from Mistral's technology page archived on February 3, 2025. ### Benchmarks Scores published by the vendor (Hugging Face model card, instruct model): | Benchmark | Mistral Small 3 | |---|---| | MMLU-Pro (5-shot CoT) | 66.3% | | GPQA Main (CoT 5-shot) | 45.3% | | HumanEval | 84.8% | | MATH | 70.6% | | IFEval | 82.9% | The announcement also cites more than 81% accuracy on MMLU (base model). ### Key takeaways An open, lightweight model, usable on a single GPU or a quantized 32 GB machine, suited to internal uses where latency and data control matter.
Links
Tags : LLM, Mistral, Small, Open Source, Apache 2.0