Llama on AMD Strix Halo

Can I run Llama 4 Scout 109B-A17B on an Ryzen AI Max+ 395 128GB?

Yes. Llama 4 Scout 109B-A17B at Q4_K_M uses 63.7 GB of the 96.0 GB available on a Ryzen AI Max+ 395 128GB, and runs at about 11.1 tokens per second. You can push the context to 180K.

Runs comfortably

63.7 GB of 96.0 GB · 66%
096 GB
Weights 61.7 GB
KV cache 1.5 GB
Runtime overhead 0.5 GB

Llama 4 Scout 109B-A17B at Q4_K_M leaves 32.3 GB spare on a Ryzen AI Max+ 395 128GB. There is room to raise the context length or move up a quantisation level.

Generation11.1tok/s
Prompt processing618tok/s
Max context180Ktokens
KV per 1K tokens0GB

Every quantisation of Llama 4 Scout 109B-A17B 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 16.00 203.0 GB 205.1 GB Won't fit 1.53
INT8 / W8A8 8.50 107.4 GB 109.4 GB Won't fit 4.86
Q8_0 (GGUF) 8.50 107.4 GB 109.4 GB Won't fit 4.86
FP8 (E4M3) 8.00 101.5 GB 103.6 GB Won't fit 5.77
Q6_K 6.56 83.2 GB 85.3 GB Runs comfortably 65K 8.38
Q5_K_M 5.67 71.9 GB 74.0 GB Runs comfortably 125K 9.61
Q5_K_S 5.52 70.0 GB 72.1 GB Runs comfortably 136K 9.86
Q4_K_M 4.85 61.7 GB 63.7 GB Runs comfortably 180K 11.1
Q4_K_S 4.58 58.3 GB 60.4 GB Runs comfortably 198K 11.7
Q4_0 4.55 58.0 GB 60.0 GB Runs comfortably 200K 11.7
AWQ 4-bit 4.25 56.8 GB 58.8 GB Runs comfortably 206K 12.0
GPTQ 4-bit 4.25 56.8 GB 58.8 GB Runs comfortably 206K 12.0
MXFP4 4.25 56.8 GB 58.8 GB Runs comfortably 206K 12.0
IQ4_XS 4.25 54.2 GB 56.3 GB Runs comfortably 220K 12.5
Q3_K_M 3.91 50.0 GB 52.0 GB Runs comfortably 242K 13.4
IQ3_M 3.70 47.4 GB 49.4 GB Runs comfortably 256K 14.1
IQ3_XXS 3.06 39.4 GB 41.5 GB Runs comfortably 299K 16.6
Q2_K 2.63 34.1 GB 36.1 GB Runs comfortably 327K 18.9
IQ2_XXS 2.06 27.0 GB 29.0 GB Runs comfortably 365K 23.1
IQ1_M 1.75 23.1 GB 25.1 GB Runs comfortably 386K 26.2

Also worth checking