Phi-3 is a family of open AI models developed by Microsoft, offering groundbreaking performance for small language models (SLMs).
Phi-3 Mini
- 3.8B parameter lightweight model
- Trained on high-quality synthetic data and filtered web data focused on reasoning
- Post-training with supervised fine-tuning and preference optimization for instruction-following and safety
- State-of-the-art performance among models <13B parameters on benchmarks for common sense, language, math, coding, reasoning
- Available in 4K and 128K context window variants
- Optimized for ONNX Runtime with GPU, CPU, mobile hardware support
Phi-3 Small and Medium
- Phi-3 Small: 7B parameters, outperforms GPT-3 175B on many benchmarks
- Phi-3 Medium: 14B parameters, outperforms Gemini 1.0 Pro
Availability
- Phi-3 Mini available on Azure AI Studio, Hugging Face, Ollama
- Phi-3 Small and Medium coming to Azure AI Model Catalog
- Optimized ONNX versions for GPU, CPU, mobile
The Phi-3 family demonstrates Microsoft's innovations in developing highly capable yet efficient and cost-effective small language models through techniques like strategic data curation and knowledge transfer.