Trained mostly on synthetic data and therefore delivers quality you would not expect from 14B: on maths and reasoning it argues with models four times its size. MIT licensed and runs on a single consumer card.
| MMLU | 84.8 | |
|---|---|---|
| GSM8K | 92.8 | |
| HumanEval | 82.6 |
Which GPU does Phi-4 14B need?
The entry configuration is 1x RTX4090, which delivers around 55 tokens per second. For more parallel requests and headroom on speed, go with 1x H100.
How much does Phi-4 14B cost?
An hour of Phi-4 14B runs from $0.54 to $3.33 depending on the GPU configuration. Billing is hourly: you pay for server time, not for tokens.
How do I call Phi-4 14B from my code?
Once deployed you get an OpenAI-compatible endpoint: point any OpenAI client at the base_url you receive and pass phi4:14b as the model name. Code written against the OpenAI API needs no changes.
Does the data stay private?
Yes. Phi-4 14B runs on a dedicated GPU server inside a private loop: requests and responses never reach the model vendor and are not used for training.
What context window and license does Phi-4 14B have?
Phi-4 14B holds 16K tokens of context per request and ships under the MIT license, so you can use it commercially on your own hosting terms.