Qwen on AMD Strix Halo

Can I run Qwen2.5 72B on an Ryzen AI Max+ 395 128GB?

Yes. Qwen2.5 72B at Q4_K_M uses 44.3 GB of the 96.0 GB available on a Ryzen AI Max+ 395 128GB, and runs at about 3.16 tokens per second. You can push the context to 128K.

Runs comfortably

44.3 GB of 96.0 GB · 46%
096 GB
Weights 41.2 GB
KV cache 2.5 GB
Runtime overhead 0.6 GB

Qwen2.5 72B at Q4_K_M leaves 51.7 GB spare on a Ryzen AI Max+ 395 128GB. There is room to raise the context length or move up a quantisation level.

Generation3.16tok/s
Prompt processing144tok/s
Max context128Ktokens
KV per 1K tokens0GB

Every quantisation of Qwen2.5 72B 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 16.00 135.4 GB 138.5 GB Won't fit 0.50
INT8 / W8A8 8.50 71.4 GB 74.5 GB Runs comfortably 77K 1.85
Q8_0 (GGUF) 8.50 71.4 GB 74.5 GB Runs comfortably 77K 1.85
FP8 (E4M3) 8.00 67.7 GB 70.8 GB Runs comfortably 89K 1.95
Q6_K 6.56 55.5 GB 58.6 GB Runs comfortably 128K 2.37
Q5_K_M 5.67 48.0 GB 51.1 GB Runs comfortably 128K 2.73
Q5_K_S 5.52 46.7 GB 49.8 GB Runs comfortably 128K 2.80
Q4_K_M 4.85 41.2 GB 44.3 GB Runs comfortably 128K 3.16
AWQ 4-bit 4.25 39.4 GB 42.5 GB Runs comfortably 128K 3.30
GPTQ 4-bit 4.25 39.4 GB 42.5 GB Runs comfortably 128K 3.30
MXFP4 4.25 39.4 GB 42.5 GB Runs comfortably 128K 3.30
Q4_K_S 4.58 39.0 GB 42.1 GB Runs comfortably 128K 3.33
Q4_0 4.55 38.8 GB 41.9 GB Runs comfortably 128K 3.35
IQ4_XS 4.25 36.3 GB 39.4 GB Runs comfortably 128K 3.57
Q3_K_M 3.91 33.6 GB 36.6 GB Runs comfortably 128K 3.86
IQ3_M 3.70 31.8 GB 34.9 GB Runs comfortably 128K 4.06
IQ3_XXS 3.06 26.6 GB 29.7 GB Runs comfortably 128K 4.82
Q2_K 2.63 23.1 GB 26.2 GB Runs comfortably 128K 5.51
IQ2_XXS 2.06 18.4 GB 21.5 GB Runs comfortably 128K 6.81
IQ1_M 1.75 15.9 GB 19.0 GB Runs comfortably 128K 7.81

Also worth checking