Microsoft Research · General LLM
Phi-3
Microsoft's compact language model family demonstrating that small models trained on high-quality data can rival much larger models in reasoning tasks.
Overview
Phi-3 is Microsoft's family of small language models that challenge the assumption that bigger is always better. The Phi-3 Mini (3.8B), Small (7B), and Medium (14B) models achieve remarkable performance by focusing on data quality over data quantity, using carefully curated training data including textbook-quality content and synthetic data. Phi-3 Mini matches or exceeds the performance of models 5-10x its size on many benchmarks, making it ideal for edge deployment and cost-sensitive applications.
Parameters
3.8B (Mini), 7B (Small), 14B (Medium)
Context Window
4K-128K tokens (variant dependent)
Training Approach
High-quality curated + synthetic data
Quantized Size
~1.8GB (Mini, 4-bit quantized)
License
MIT
Capabilities
Strong reasoning in a compact model footprint
Code generation competitive with larger models
On-device and edge deployment capability
Efficient inference on consumer hardware
Mathematical problem solving
Use Cases
Deploying AI on mobile devices and edge hardware
Running efficient AI inference on laptops without a GPU
Building cost-effective AI applications at scale
Embedding AI capabilities in resource-constrained environments
Pros
- +Exceptional performance-to-size ratio
- +Small enough for mobile and edge deployment
- +MIT license enables unrestricted commercial use
- +Demonstrates the power of data quality over quantity
Cons
- -Smaller knowledge base than frontier-scale models
- -Limited multilingual capabilities compared to larger models
- -Context window is shorter on the Mini variant
- -May struggle with complex multi-step reasoning tasks
Pricing
Free and open-source under MIT license. Runs on consumer hardware. Azure AI offers hosted inference at competitive per-token pricing.