Yes, just — 0.2 GB spare
Phi-4 14B at Q4_K_M fits your GeForce RTX 2060 12 GB entirely on the GPU at 8K context, at an estimated 25 tokens per second. Past 8K the KV cache pushes it over — quantise the cache to q8_0, or step down a quantisation, to go longer.
Fully 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.
The VRAM budget
weights 8.3 GB
Weights 8.3 GB
KV cache @ 8K 1.56 GB
Runtime overhead 0.6 GB
Free 0.2 GB of 10.6 GB
Quantisation ladder
| Quant | Weights | Total @ 8K | Max context | Tok/s | Quality | Fit |
|---|---|---|---|---|---|---|
| F16 | 27.4 GB | 29.5 GB | — | ~1.8 | Reference | 18.9 GB over |
| Q8_0 | 14.5 GB | 16.7 GB | — | ~4.8 | −0.1% ppl | 6.1 GB over |
| Q6_K | 11.2 GB | 13.4 GB | — | ~8.5 | −0.4% ppl | 2.8 GB over |
| Q5_K_M | 9.7 GB | 11.9 GB | 1K | ~13 | −0.8% ppl | 1.3 GB over |
| Q4_K_M | 8.3 GB | 10.4 GB | 8K | 25 | −1.9% ppl | Recommended |
| Q3_K_M | 6.7 GB | 8.9 GB | 16K | 30 | −5.4% ppl | Fits |
| Q2_K | 5.7 GB | 7.9 GB | 16K | 35 | −15% ppl | Fits |
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 99
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.
03Only 0.2 GB is spare, so a long system prompt can still push it over. Its real ceiling here is 8K context.