Phi-4 Reasoning Vision
Phi-4 Reasoning
Fifteen billion parameters that reason like something several times the size.
Updated
The read
Microsoft's Phi line exists to prove that data quality beats parameter count, and this release is the strongest argument yet: a small MIT-licensed model that handles multi-step reasoning and reads images, on hardware a single developer can afford. It is the obvious choice for an on-device assistant or an edge deployment that still has to think. World knowledge is thin, because there is nowhere to put it.
Where it shines
- Reasoning well above what its size would predict
- MIT licence, so it ships inside commercial products freely
- Runs on a single consumer GPU with room to spare
Worth knowing first
- Thin world knowledge: it reasons better than it recalls
- Long-context work is not what it was built for
Index score
76.8
Provisional. This score is read from the product's public capability, not from the one-prompt rebuild the fully reviewed entries went through, so treat it as a placing rather than a verdict.
- Rank
- 59 of 73
- Builds
- LLMs
- Category
- Small and fast
- Output
- MIT-licensed weights, small and multimodal
- Pricing
- Free tier
- Best for
- On-device reasoning without a datacentre behind it
Scores are our own editorial judgement, weighted craft 55, speed 10, control 15 and value 20.
Also worth a look
Full index- 11Claude Haiku 5.5Claude Haiku 5.5ProvisionalA small model priced like a rounding error that scores where mid-tier models did last quarter.Small and fastPaid only87.9
- 22GPT-6 LunaGPT-6 LunaProvisionalThe cheap end of GPT-6, and the model free ChatGPT users now talk to.Small and fastPaid only84.9
- 30GPT-5.6 LunaGPT-5.6 LunaProvisionalStill cheap and still capable, but GPT-6 Luna costs half as much.Small and fastPaid only83.0
