Microsoft on AMD Strix Halo

Can I run Phi-4 14B on an Ryzen AI Max+ 395 128GB?

Yes. Phi-4 14B at Q4_K_M uses 10.5 GB of the 96.0 GB available on a Ryzen AI Max+ 395 128GB, and runs at about 14.6 tokens per second. You can push the context to 16K.

Runs comfortably

10.5 GB of 96.0 GB · 11%
096 GB
Weights 8.4 GB
KV cache 1.6 GB
Runtime overhead 0.5 GB

Phi-4 14B at Q4_K_M leaves 85.5 GB spare on a Ryzen AI Max+ 395 128GB. There is room to raise the context length or move up a quantisation level.

Generation14.6tok/s
Prompt processing714tok/s
Max context16Ktokens
KV per 1K tokens0GB

Every quantisation of Phi-4 14B on a Ryzen AI Max+ 395 128GB

Highlighted row is the highest quality that still fits at 8K context.

QuantisationbpwWeights TotalVerdictMax ctxtok/s
FP16 / BF16 16.00 27.4 GB 29.5 GB Runs comfortably 16K 4.77
INT8 / W8A8 8.50 14.3 GB 16.4 GB Runs comfortably 16K 8.89
Q8_0 (GGUF) 8.50 14.3 GB 16.4 GB Runs comfortably 16K 8.89
FP8 (E4M3) 8.00 13.7 GB 15.8 GB Runs comfortably 16K 9.27
Q6_K 6.56 11.2 GB 13.3 GB Runs comfortably 16K 11.2
Q5_K_M 5.67 9.7 GB 11.8 GB Runs comfortably 16K 12.8
Q5_K_S 5.52 9.4 GB 11.5 GB Runs comfortably 16K 13.1
AWQ 4-bit 4.25 8.7 GB 10.8 GB Runs comfortably 16K 14.2
GPTQ 4-bit 4.25 8.7 GB 10.8 GB Runs comfortably 16K 14.2
MXFP4 4.25 8.7 GB 10.8 GB Runs comfortably 16K 14.2
Q4_K_M 4.85 8.4 GB 10.5 GB Runs comfortably 16K 14.6
Q4_K_S 4.58 7.9 GB 10.0 GB Runs comfortably 16K 15.3
Q4_0 4.55 7.9 GB 10.0 GB Runs comfortably 16K 15.4
IQ4_XS 4.25 7.4 GB 9.5 GB Runs comfortably 16K 16.3
Q3_K_M 3.91 6.9 GB 9.0 GB Runs comfortably 16K 17.5
IQ3_M 3.70 6.5 GB 8.6 GB Runs comfortably 16K 18.3
IQ3_XXS 3.06 5.5 GB 7.6 GB Runs comfortably 16K 21.2
Q2_K 2.63 4.8 GB 6.9 GB Runs comfortably 16K 23.8
IQ2_XXS 2.06 3.9 GB 6.0 GB Runs comfortably 16K 28.3
IQ1_M 1.75 3.4 GB 5.5 GB Runs comfortably 16K 31.6

Also worth checking

Phi-4 14B on every GPU
Full hardware tableEverything that runs on a Ryzen AI Max+ 395 128GB
Full model tableWhy the KV cache is the part that bites
Guide