Microsoft · MIT
Phi-4
Microsoft’s small models trained on textbook-style data. Good at maths for their size.
How much memory it needs
The model file, plus memory for the conversation, plus about 1 GB for the app. Longer conversations need more.
| Size | Version | File | At 8K | At 32K | At 128K |
|---|---|---|---|---|---|
| 3.8B | Q4_K_M | 2.5 GB | 4.6 GB | 7.8 GB | 20.6 GB |
| Q8_0 | 4.1 GB | 6.2 GB | 9.4 GB | 22.2 GB | |
| 15B | Q4_K_M | 8.9 GB | 11.6 GB | — | — |
| Q8_0 | 15.6 GB | 18.3 GB | — | — |
Which computers run it
The best size for each machine at an 8K-token conversation. Speeds are estimates from memory bandwidth.
| Computer | Best version | Fit | Speed |
|---|---|---|---|
| Laptop or PC, no graphics card, 32 GB | 3.8B Q8_0 | Runs great | 10–17tok/sReadable |
| RTX 4060 · 8 GB | 3.8B Q8_0 | Runs well | 30–50tok/sFast |
| RTX 3060 · 12 GB | 3.8B Q8_0 | Runs great | 40–65tok/sFast |
| RTX 5060 Ti · 16 GB | 15B Q4_K_M | Runs great | 25–40tok/sSmooth |
| RTX 4090 · 24 GB | 15B Q8_0 | Runs well | 35–55tok/sFast |
| RTX 5090 · 32 GB | 15B Q8_0 | Runs great | 60–95tok/sFast |
| Mac with M4 · 16 GB | 3.8B Q8_0 | Runs great | 13–25tok/sSmooth |
| Mac with M4 Pro · 48 GB | 15B Q8_0 | Runs great | 9–16tok/sReadable |
| Mac with M5 Max · 128 GB | 15B Q8_0 | Runs great | 20–35tok/sSmooth |
| Ryzen AI Max+ 395 · 128 GB | 15B Q8_0 | Runs great | 9–14tok/sReadable |
| DGX Spark · 128 GB | 15B Q8_0 | Runs great | 9.5–15tok/sReadable |
| RTX PRO 6000 · 96 GB | 15B Q8_0 | Runs great | 60–95tok/sFast |
| Mac Studio with M5 Ultra · 256 GB | 15B Q8_0 | Runs great | 30–50tok/sFast |
| Mac Studio with M3 Ultra · 512 GB | 15B Q8_0 | Runs great | 20–35tok/sSmooth |
What it is good and bad at
Good at
- Maths and reasoning for a small model
- MIT licence
Not good at
- Phi-4 14B has only a 16K context
- Text only
Start it
Install Ollama, then run one of these in a terminal. LM Studio users can search the same name in the app.
ollama run phi4-mini:3.8bollama run phi4:14b