Skip to content

Phi 4 Reasoning VRAM requirements

How much VRAM Phi 4 Reasoning needs
4-bit weights: 8.3 GBWith 8K context: 10.6 GBWith 32K context: 15.3 GB8-bit, 8K context: 16.9 GB
Parameters
14.7B
Longest context
32K
Released
Apr 2025
License
MIT
Strongest at
Not rated

How much VRAM Phi 4 Reasoning needs

Memory Phi 4 Reasoning needs at each precision and context length
PrecisionWeightsWith 8KWith 32K
16-bit27.3 GB29.7 GB34.4 GB
8-bit14.5 GB16.9 GB21.6 GB
6-bit11.2 GB13.6 GB18.2 GB
5-bit9.7 GB12.1 GB16.8 GB
4-bit8.3 GB10.6 GB15.3 GB
3-bit6.7 GB9.0 GB13.7 GB
2-bit5.1 GB7.5 GB12.2 GB
1.6-bit2.8 GB5.2 GB9.9 GB

One conversation, 16-bit context cache, runtime overhead included.

GPUs and Macs that run Phi 4 Reasoning

The best way Phi 4 Reasoning runs on common GPUs, Macs and AI boxes
MachineBest fitContextSpeedFirst replyReport
RTX 4060Too bigPDF: Phi 4 Reasoning on RTX 4060 report
RTX 5060Too bigPDF: Phi 4 Reasoning on RTX 5060 report
RTX 3070Too bigPDF: Phi 4 Reasoning on RTX 3070 report
RTX 3080 10 GBToo bigPDF: Phi 4 Reasoning on RTX 3080 10 GB report
RTX 3060 12 GB3-bit20K35 tok/s2.4 sPDF: Phi 4 Reasoning on RTX 3060 12 GB report
Arc B5803-bit20K31 tok/s1.6 sPDF: Phi 4 Reasoning on Arc B580 report
RTX 40703-bit20K48 tok/s1.0 sPDF: Phi 4 Reasoning on RTX 4070 report
RTX 4070 Super3-bit20K48 tok/s0.9 sPDF: Phi 4 Reasoning on RTX 4070 Super report
RTX 50703-bit20K63 tok/s1.0 sPDF: Phi 4 Reasoning on RTX 5070 report
RTX 4060 Ti 16 GB6-bit16K17 tok/s1.4 sPDF: Phi 4 Reasoning on RTX 4060 Ti 16 GB report
RTX 5060 Ti 16 GB6-bit16K26 tok/s1.3 sPDF: Phi 4 Reasoning on RTX 5060 Ti 16 GB report
Arc A770 16 GB6-bit16K23 tok/s1.5 sPDF: Phi 4 Reasoning on Arc A770 16 GB report
Radeon RX 9070 XT6-bit16K31 tok/s0.8 sPDF: Phi 4 Reasoning on Radeon RX 9070 XT report
RTX 4070 Ti Super6-bit16K39 tok/s0.7 sPDF: Phi 4 Reasoning on RTX 4070 Ti Super report
RTX 4080 Super6-bit16K42 tok/s0.6 sPDF: Phi 4 Reasoning on RTX 4080 Super report
RTX 5070 Ti6-bit16K51 tok/s0.7 sPDF: Phi 4 Reasoning on RTX 5070 Ti report
RTX 50806-bit16K54 tok/s0.5 sPDF: Phi 4 Reasoning on RTX 5080 report
Radeon RX 7900 XTX8-bit32K36 tok/s1.3 sPDF: Phi 4 Reasoning on Radeon RX 7900 XTX report
RTX 30908-bit32K41 tok/s0.9 sPDF: Phi 4 Reasoning on RTX 3090 report
RTX 40908-bit32K44 tok/s0.4 sPDF: Phi 4 Reasoning on RTX 4090 report
Radeon AI PRO R97008-bit32K24 tok/s0.9 sPDF: Phi 4 Reasoning on Radeon AI PRO R9700 report
RTX 50908-bit32K76 tok/s0.3 sPDF: Phi 4 Reasoning on RTX 5090 report
2× RTX 30908-bit32K41 tok/s0.5 sPDF: Phi 4 Reasoning on 2× RTX 3090 report
2× RTX 40908-bit32K44 tok/s0.2 sPDF: Phi 4 Reasoning on 2× RTX 4090 report
RTX PRO 5000 Blackwell 48 GB8-bit32K58 tok/s0.4 sPDF: Phi 4 Reasoning on RTX PRO 5000 Blackwell 48 GB report
RTX 6000 Ada8-bit32K42 tok/s0.3 sPDF: Phi 4 Reasoning on RTX 6000 Ada report
L40S8-bit32K38 tok/s0.5 sPDF: Phi 4 Reasoning on L40S report
2× RTX 50908-bit32K76 tok/s0.2 sPDF: Phi 4 Reasoning on 2× RTX 5090 report
4× RTX 30908-bit32K41 tok/s0.3 sPDF: Phi 4 Reasoning on 4× RTX 3090 report
RTX PRO 6000 Blackwell8-bit32K76 tok/s0.2 sPDF: Phi 4 Reasoning on RTX PRO 6000 Blackwell report
Mac mini M4 16 GBToo bigPDF: Phi 4 Reasoning on Mac mini M4 16 GB report
Mac mini M4 24 GB6-bit16K8.3 tok/s19 sPDF: Phi 4 Reasoning on Mac mini M4 24 GB report
Mac mini M4 32 GB8-bit24K6.4 tok/s19 sPDF: Phi 4 Reasoning on Mac mini M4 32 GB report
Mac mini M4 Pro 24 GB6-bit16K18 tok/s9.6 sPDF: Phi 4 Reasoning on Mac mini M4 Pro 24 GB report
Mac mini M4 Pro 48 GB8-bit32K14 tok/s9.6 sPDF: Phi 4 Reasoning on Mac mini M4 Pro 48 GB report
Mac mini M4 Pro 64 GB8-bit32K14 tok/s9.6 sPDF: Phi 4 Reasoning on Mac mini M4 Pro 64 GB report
M4 Max 128 GB8-bit32K25 tok/s4.8 sPDF: Phi 4 Reasoning on M4 Max 128 GB report
M5 Max 128 GB8-bit32K28 tok/s1.4 sPDF: Phi 4 Reasoning on M5 Max 128 GB report
M3 Ultra 96 GB8-bit32K32 tok/s2.7 sPDF: Phi 4 Reasoning on M3 Ultra 96 GB report
M3 Ultra 256 GB8-bit32K32 tok/s2.7 sPDF: Phi 4 Reasoning on M3 Ultra 256 GB report
M3 Ultra 512 GB8-bit32K32 tok/s2.7 sPDF: Phi 4 Reasoning on M3 Ultra 512 GB report
M5 Ultra 512 GB8-bit32K45 tok/s0.8 sPDF: Phi 4 Reasoning on M5 Ultra 512 GB report
Ryzen AI Max+ 395 64 GB8-bit32K10 tok/s2.8 sPDF: Phi 4 Reasoning on Ryzen AI Max+ 395 64 GB report
Ryzen AI Max+ 395 128 GB8-bit32K10 tok/s2.8 sPDF: Phi 4 Reasoning on Ryzen AI Max+ 395 128 GB report
RTX Spark8-bit32K14 tok/s1.0 sPDF: Phi 4 Reasoning on RTX Spark report
DGX Spark8-bit32K13 tok/s1.0 sPDF: Phi 4 Reasoning on DGX Spark report
Dual DGX Spark8-bit32K13 tok/s0.6 sPDF: Phi 4 Reasoning on Dual DGX Spark report
DGX Station GB3008-bit32K237 tok/s0.1 sPDF: Phi 4 Reasoning on DGX Station GB300 report
2× RTX PRO 6000 Blackwell8-bit32K76 tok/s0.1 sPDF: Phi 4 Reasoning on 2× RTX PRO 6000 Blackwell report
4× RTX PRO 6000 Blackwell8-bit32K76 tok/s0.1 sPDF: Phi 4 Reasoning on 4× RTX PRO 6000 Blackwell report
8× RTX PRO 6000 Blackwell8-bit32K76 tok/s0.1 sPDF: Phi 4 Reasoning on 8× RTX PRO 6000 Blackwell report
A100 80 GB SXM8-bit32K85 tok/s0.6 sPDF: Phi 4 Reasoning on A100 80 GB SXM report
H100 SXM8-bit32K131 tok/s0.2 sPDF: Phi 4 Reasoning on H100 SXM report
4× H200 NVL8-bit32K176 tok/s0.1 sPDF: Phi 4 Reasoning on 4× H200 NVL report
8× H200 NVL8-bit32K176 tok/s0.1 sPDF: Phi 4 Reasoning on 8× H200 NVL report

Best fit is the highest quality, up to 8-bit, that leaves 16K of context. First reply is the wait for a 2K prompt. Speeds are rough, plus or minus 25%.

How to run it

llama.cpp
llama-server -hf bartowski/microsoft_Phi-4-reasoning-GGUF
vLLM
vllm serve microsoft/Phi-4-reasoning
MLX (Mac)
mlx_lm.chat --model mlx-community/Phi-4-reasoning-4bit

Questions

How much VRAM does Phi 4 Reasoning need?

Phi 4 Reasoning needs 8.3 GB for its weights at 4-bit, 10.6 GB with an 8K context and 15.3 GB with 32K. At 8-bit, 16.9 GB.

What is the minimum VRAM for Phi 4 Reasoning?

About 9.0 GB at 3-bit with an 8K context, the lowest precision we recommend for a model this size. The smallest machine here that runs it without spilling is the RTX 3060 12 GB.

How big is Phi 4 Reasoning?

14.7B parameters: 27.3 GB of weights at 16-bit, 8.3 GB at 4-bit.

Can Phi 4 Reasoning run on an RTX 4090?

Yes. At 8-bit it holds 32K of context and writes about 44 tokens a second.

More models