Phi-4-reasoning
Open Weights
Microsoft
Microsoft's open-weight 14B reasoning model (April 2025, MIT license), fine-tuned from Phi-4 for chain-of-thought math and scientific reasoning. It scores 75.3% on AIME 2024 (81.3% for the reasoning-plus variant), beating much larger distilled models, with a 32K context window stable to 64K.
Strengths
- Open-weight 14B reasoning model under MIT license
- 75.3% AIME 2024 - 81.3% for the reasoning-plus variant
- Beats DeepSeek-R1-Distill-70B on AIME despite far fewer parameters
- 32K context stable to 64K in experiments
- Strong math, science, coding, and planning reasoning
Caveats
- Reasoning tuning trades some general-purpose breadth
- Text-only (no vision or audio input)
- 32K context smaller than frontier reasoning models
- Vendor
- Microsoft
- Context Window
- 32,000 tokens
- License
- MIT
- Release Date
- 2025-04-30
- Modalities
- text
| Benchmark | Score | Max |
|---|---|---|
| AIME 2024 | 75.3 | 100.0 |
Capabilities
Vision
Audio
Video
Tool Use
Resources
Pricing
Open weights under MIT license - free to download and self-host; 14B runs on consumer/prosumer GPUs.