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 8.9 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.

Hardware requirements What hardware do I need to run Phi-4 14B? →

At Q4_K_M with an 8K context: a 16 GB GPU (RTX 5080) or a 24 GB Mac (M4 Pro). Minimum, recommended, high-quality and long-context tiers, with all 114 devices checked at every quantisation.

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

QuantWeightsTotal @ 8KMax contextTok/sQualityFit
F16 27.4 GB 29.5 GB — ~1.6 Reference 22.5 GB over
Q8_0 14.5 GB 16.7 GB — ~3.5 −0.1% ppl 9.7 GB over
Q6_K 11.2 GB 13.4 GB — ~5.2 −0.4% ppl 6.4 GB over
Q5_K_M 9.7 GB 11.9 GB — ~6.6 −0.8% ppl 4.9 GB over
Q4_K_M 8.3 GB 10.4 GB — ~8.9 −1.9% ppl 3.4 GB over
Q3_K_M 6.7 GB 8.9 GB — ~15 −5.4% ppl 1.9 GB over
Q2_K 5.7 GB 7.9 GB 3K ~24 −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

terminal
$ 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.
See all models for this rig Compare with another model

Related models on a GeForce RTX 5060

ModelTotalTok/s est.
Phi-4-mini 3.8B
3.84B · same family
Q4_K_M · 3.8 GB 126 Runs great
Qwen3 14B
14.8B · similar size
Q4_K_M · 10.2 GB ~9.4 CPU offload
DeepSeek-R1-Distill-Qwen 14B
14.8B · similar size
Q4_K_M · 10.4 GB ~8.9 CPU offload
Qwen2.5-Coder 14B
14.8B · similar size
Q4_K_M · 10.4 GB ~8.9 CPU offload
Ministral 3 14B
13.9B · similar size
Q4_K_M · 9.7 GB ~11 CPU offload
Gemma 3 12B
12.2B · similar size
Q4_K_M · 8.3 GB ~18 CPU offload
Mistral NeMo 12B
12.2B · similar size
Q4_K_M · 8.7 GB ~15 CPU offload

Every model against the GeForce RTX 5060 →

Phi-4 14B on other hardware

DeviceTotalTok/s est.
GeForce RTX 4060 Laptop
7.0 GB usable · 256 GB/s
10.4 GB ~8.0 CPU offload
GeForce RTX 3060 12 GB
10.6 GB usable · 360 GB/s
10.4 GB 26 Tight fit
GeForce RTX 5080
14.4 GB usable · 960 GB/s
10.4 GB 70 Runs great
M4 · 16 GB
10.7 GB usable · 120 GB/s
10.4 GB 8.1 Tight fit
GeForce RTX 4090
22.4 GB usable · 1008 GB/s
10.4 GB 74 Runs great
Radeon RX 7900 XTX
22.4 GB usable · 960 GB/s
10.4 GB 70 Runs great
M4 Pro · 24 GB
18.0 GB usable · 273 GB/s
10.4 GB 18 Runs great
M3 Max · 64 GB
48.0 GB usable · 400 GB/s
10.4 GB 27 Runs great

All 114 devices, by tier →