Mistral Small 3 is a 24-billion-parameter open-source large language model released by Mistral AI, a French AI startup. The model represents a breakthrough in efficiency, achieving performance comparable to models 2-3 times its size.
- Parameters: 24 billion
- Model Type: Transformer-based language model
- Efficiency: Optimized architecture for faster inference
- Training: Advanced training on high-quality multilingual data
- Performance on par with 70B models like LLaMA 3.3
- 3x faster inference on same hardware compared to larger models
- Strong multilingual capabilities
- Excellent reasoning abilities
- Optimized for efficiency without sacrificing quality
- Superior instruction following
- Enhanced coding capabilities
- Matches or exceeds LLaMA 70B on many benchmarks
- Strong MMLU and reasoning scores
- Competitive multilingual performance
- Excellent code generation results
- High efficiency-to-performance ratio
- Self-hosting on consumer and enterprise hardware
- Cloud deployment options
- Optimized for GPU inference
- Compatible with standard ML frameworks
- Support for quantization
- Enterprise applications requiring efficient inference
- Multilingual conversational AI
- Code generation and assistance
- Content creation and analysis
- Research and development
- Edge deployment scenarios
Released under Apache 2.0 license, a welcome change from earlier Mistral releases that used more restrictive research licenses. This allows full commercial use without restrictions.