| Phi-4 |
Jun 2025 |
14 |
Advanced reasoning, low latency, trained on synthetic and academic data |
| Phi-4-mini-instruct |
Feb 2025 |
3.8 |
Multilingual, reasoning, math, supports function calling, edge deployable |
| Phi-4-multimodal |
Feb 2025 |
5.6 |
Multimodal (text, vision, audio), strong reasoning, edge deployable |
| Phi-3-mini |
Nov 2024 |
3.8 |
4K & 128K context variants, trained on 3.3T tokens, strong benchmarks |
| Phi-3-small |
Nov 2024 |
7 |
8K context, optimized memory via Grouped Query Attention |
| Phi-3-medium |
Nov 2024 |
14 |
Larger variant, in training phase, designed for general-purpose AI |
| Phi-2 |
2023 |
2.7 |
Predecessor to Phi-3, focused on compact reasoning models |
| Phi-1.5 |
2023 |
1.3 |
Early small language model, used for foundational research |