Skip to content
localmodel.run

Text model · Phi-4

Phi-4-reasoning: RAM and VRAM requirements

Phi-4-reasoning needs about 10.1 GB to run at Q4_K_M (the Q4_K_M GGUF file is ~8.43 GB to download; KV cache and overhead add the rest), or about 16.2 GB at Q8_0. The lightest hardware that runs it is Nvidia GeForce RTX 3060 (12GB).

Phi-4 family · 14B params · released Apr 2025.

Shopping for hardware? See what runs Phi-4-reasoning →

License MIT · Commercial OK ↓ 23.7K/mo ♥ 228 on HuggingFace
Q4_K_M GGUF
8.43 GB
Q8_0 GGUF
14.51 GB
Memory @ Q4 (4k)
~10.1 GB
Context
32 k

Q4_K_M and Q8_0 are GGUF download (file) sizes. Memory @ Q4 includes the KV cache and overhead, so it runs larger than the file on disk.

Will it run on your device?

Phi-4-reasoning runs on 29 of 40 tracked devices at Q4_K_M.

22 run well 7 tight fit 11 too small
Fit check Q4_K_M
Yes, but tight fast
needs 10.1 GB usable 10.5 GB
See the full breakdown
$ollama run phi4-reasoning:14b

Memory breakdown

Weights (Q4_K_M)8.43 GB
+
KV cache (4k)0.9 GB
+
Overhead0.8 GB
=
Total10.1 GB

How context length changes it

4k context ~10.1 GB 32k context ~16.4 GB 128k context ~37.9 GB

Longer context grows the KV cache quickly: Phi-4-reasoning needs ~10.1 GB at 4k but ~37.9 GB at 128k, which can push it past a device that fits it at a short context.

Speed drops too: every token re-reads the KV cache, so at 128k context Phi-4-reasoning generates at roughly ~23% of its short-context speed. How this is estimated.

Quantization sizes

GGUF quantson disk
Quantization Size on disk
Q2_K 5.9 GB est
Q3_K_M 6.8 GB est
Q4_K_M (default) 8.43 GB
Q5_K_M 10 GB est
Q6_K 11.5 GB est
Q8_0 14.51 GB
FP16 29.32 GB

Lower quant = smaller and faster, slightly lower quality. Q4_K_M is the common default.

Run it

Ollama
$ ollama run phi4-reasoning:14b
llama.cpp
$ llama-cli -hf bartowski/microsoft_Phi-4-reasoning-GGUF:Q4_K_M
LM Studio
$ lms get bartowski/microsoft_Phi-4-reasoning-GGUF

Which devices can run Phi-4-reasoning?

Apple Silicon Macs

RAM-only laptops

iPhone & iPad

Android

NVIDIA GPUs

AMD GPUs

Same job, different size

Pick by what fits your memory: step down to free up VRAM, or step up if you have headroom.

Next size down Mistral Nemo 12B 12.2B Next size up DeepSeek-V2-Lite 16B

Similar models

Head-to-head

FAQ

How much VRAM or RAM does Phi-4-reasoning need?

At Q4_K_M, Phi-4-reasoning needs about 10.1 GB (weights ~8.43 GB + KV cache + overhead) at a 4k context. At Q8_0 budget ~16.2 GB.

What is the Q4_K_M GGUF file size of Phi-4-reasoning?

The Q4_K_M GGUF file is about 8.43 GB to download, and the Q8_0 GGUF is about 14.51 GB. That is the weights file on disk; to run it you also need room for the KV cache and overhead, so budget ~10.1 GB of memory at Q4_K_M.

Can Phi-4-reasoning run on a laptop?

Phi-4-reasoning is large; you need a 24 GB+ GPU or a 32-48 GB Mac at Q4_K_M.

Can I use Phi-4-reasoning commercially?

Yes. Phi-4-reasoning is licensed MIT, which permits commercial use.

Understand the numbers

Short guides to the ideas behind Phi-4-reasoning's memory and quant figures.

Microsoft's reasoning-tuned Phi-4. Strong math and logic for a 14B; runs on a 16GB machine at Q4_K_M.

Sources

Last validated 2026-08-03. Memory figures are estimates. See methodology.