Microsoft: Phi 4
Microsoft · microsoft/phi-4
Runs in the cloud on our API and on your own machine — phi4:14b on the local shelf, free and offline.
open weights
[Microsoft Research](/microsoft) Phi-4 is designed to perform well in complex reasoning tasks and can operate efficiently in situations with limited memory or where quick responses are needed. At 14 billion...
What it costs
Final AI-token prices, per million tokens. Bring your own provider key in the app to pay provider rates directly.
What it can do
Benchmarks
| Benchmark | What it measures | Score |
|---|---|---|
| Epoch Capabilities Index | one score summarising the benchmarks below — an index, not a percentage | 130.4 |
| GPQA Diamond | graduate-level science questions | 56.1% ± 2.6 |
| AIME | competition mathematics | 13.8% ± 3.7 |
| MATH level 5 | the hardest tier of the MATH problem set | 64.9% ± 1.1 |
Scores by Epoch AI, CC BY 4.0. The ± is the margin of error: models inside each other's margins are tied, not ranked. Epoch's notes on this row: source “epoch”, accessibility “Open weights (unrestricted)”.
Dates
Using it
In the 00 app — pick it per agent, and per tier, then build your own agents on it. Get 00.
As a drop-in API — the same model at an OpenAI-compatible endpoint. Two settings:
- Create a workspace API key in the Overblast console — usage bills that workspace's AI tokens, at the prices on this page.
- Point your base URL at
https://brain.deployd.network/ai/v1. Chat completions with streaming, embeddings,/images,/audio/transcriptionsand/videos.
curl https://brain.deployd.network/ai/v1/chat/completions \
-H "Authorization: Bearer ob_live_YOUR_KEY" \
-H "Content-Type: application/json" \
-d '{
"model": "microsoft/phi-4",
"messages": [{ "role": "user", "content": "Hello" }]
}'
← all models · compare it with another · calling image, video and speech models
Subscribe to new models