Qwen on AMD Strix Halo

Can I run Qwen3 235B-A22B on an Ryzen AI Max+ 395 128GB?

Not at Q4_K_M — it needs 134.8 GB against 96.0 GB available. Drop to IQ3_XXS and it fits, at about 12.1 tokens per second.

Won't fit

134.8 GB of 96.0 GB · 140%
0140 GB
Weights 132.8 GB
KV cache 1.5 GB
Runtime overhead 0.5 GB
Over the limit 38.8 GB

Short by 38.8 GB. You can run it with 66 of 94 layers on the Ryzen AI Max+ 395 128GB and the rest in system RAM, at roughly 4.85 tok/s — usable for batch work, painful for chat. A smaller quantisation or a shorter context is usually the better trade.

Generation4.85tok/s
Prompt processing474tok/s
Max context0tokens
KV per 1K tokens0GB

Every quantisation of Qwen3 235B-A22B 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 16.00 437.9 GB 439.9 GB Won't fit 0.93
INT8 / W8A8 8.50 232.3 GB 234.3 GB Won't fit 2.05
Q8_0 (GGUF) 8.50 232.3 GB 234.3 GB Won't fit 2.05
FP8 (E4M3) 8.00 218.9 GB 220.9 GB Won't fit 2.21
Q6_K 6.56 179.5 GB 181.5 GB Won't fit 2.95
Q5_K_M 5.67 155.2 GB 157.2 GB Won't fit 3.67
Q5_K_S 5.52 151.1 GB 153.1 GB Won't fit 3.87
Q4_K_M 4.85 132.8 GB 134.8 GB Won't fit 4.85
Q4_K_S 4.58 125.5 GB 127.5 GB Won't fit 5.42
Q4_0 4.55 124.7 GB 126.7 GB Won't fit 5.46
AWQ 4-bit 4.25 118.0 GB 120.0 GB Won't fit 6.10
GPTQ 4-bit 4.25 118.0 GB 120.0 GB Won't fit 6.10
MXFP4 4.25 118.0 GB 120.0 GB Won't fit 6.10
IQ4_XS 4.25 116.5 GB 118.5 GB Won't fit 6.28
Q3_K_M 3.91 107.2 GB 109.2 GB Won't fit 7.64
IQ3_M 3.70 101.5 GB 103.5 GB Won't fit 8.83
IQ3_XXS 3.06 84.1 GB 86.1 GB Runs comfortably 62K 12.1
Q2_K 2.63 72.4 GB 74.4 GB Runs comfortably 126K 13.6
IQ2_XXS 2.06 56.9 GB 58.9 GB Runs comfortably 128K 16.5
IQ1_M 1.75 48.4 GB 50.4 GB Runs comfortably 128K 18.6

Also worth checking