Only with CPU offload
Phi-4 14B at Q4_K_M needs 10.4 GB but only 7.0 GB is addressable, so about 41% of the layers would stream from system RAM at roughly 60 GB/s. Expect around 9.3 tokens per second — usable for batch work, painful for chat.
59% on GPU
8K context
Q4_K_M · 8.3 GB
MIT
Released Dec 2024
Trained heavily on synthetic reasoning data. Short 16K window is its main limitation.
What hardware do I need for Phi-4 14B? →
Fits instead: Phi-4-mini 3.8B (3.8 GB)
The VRAM budget
weights 8.3 GB
Weights 8.3 GB
KV cache @ 8K 1.56 GB
Runtime overhead 0.6 GB
Over budget 3.4 GB past 7.0 GB
Quantisation ladder
| Quant | Weights | Total @ 8K | Max context | Tok/s | Quality | Fit |
|---|---|---|---|---|---|---|
| F16 | 27.4 GB | 29.5 GB | — | ~1.6 | Reference | 22.5 GB over |
| Q8_0 | 14.5 GB | 16.7 GB | — | ~3.6 | −0.1% ppl | 9.7 GB over |
| Q6_K | 11.2 GB | 13.4 GB | — | ~5.3 | −0.4% ppl | 6.4 GB over |
| Q5_K_M | 9.7 GB | 11.9 GB | — | ~6.8 | −0.8% ppl | 4.9 GB over |
| Q4_K_M | 8.3 GB | 10.4 GB | — | ~9.3 | −1.9% ppl | 3.4 GB over |
| Q3_K_M | 6.7 GB | 8.9 GB | — | ~16 | −5.4% ppl | 1.9 GB over |
| Q2_K | 5.7 GB | 7.9 GB | 3K | ~26 | −15% ppl | 0.9 GB over |
Quality is the published perplexity delta against f16 weights. Max context assumes an f16 KV cache; q8_0 roughly doubles it.
How to run it
$ llama-server \
-hf microsoft/phi-4:Q4_K_M \
-c 8192 -ngl 23
The engine underneath most of the others. Every knob is exposed. More on llama.cpp.
01Download is 8.3 GB. Keep it on an SSD — a first load off a spinning disk takes minutes.
02Close anything else holding VRAM. A browser with hardware acceleration can sit on 1–2 GB.
03If it falls back to CPU silently, drop the context first, then step down a quantisation.