Llama on AMD Strix Halo

Can I run Llama 3.2 1B on an Ryzen AI Max+ 395 128GB?

Yes. Llama 3.2 1B at Q4_K_M uses 1.5 GB of the 96.0 GB available on a Ryzen AI Max+ 395 128GB, and runs at about 157 tokens per second. You can push the context to 128K.

Runs comfortably

1.5 GB of 96.0 GB · 2%
096 GB
Weights 0.7 GB
KV cache 0.3 GB
Runtime overhead 0.5 GB

Llama 3.2 1B at Q4_K_M leaves 94.5 GB spare on a Ryzen AI Max+ 395 128GB. There is room to raise the context length or move up a quantisation level.

Generation157tok/s
Prompt processing8468tok/s
Max context128Ktokens
KV per 1K tokens0GB

Every quantisation of Llama 3.2 1B 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 2.3 GB 3.0 GB Runs comfortably 128K 54.9
INT8 / W8A8 8.50 1.2 GB 1.9 GB Runs comfortably 128K 103
Q8_0 (GGUF) 8.50 1.2 GB 1.9 GB Runs comfortably 128K 103
FP8 (E4M3) 8.00 1.2 GB 1.9 GB Runs comfortably 128K 104
AWQ 4-bit 4.25 1.0 GB 1.7 GB Runs comfortably 128K 121
GPTQ 4-bit 4.25 1.0 GB 1.7 GB Runs comfortably 128K 121
MXFP4 4.25 1.0 GB 1.7 GB Runs comfortably 128K 121
Q6_K 6.56 0.9 GB 1.7 GB Runs comfortably 128K 124
Q5_K_M 5.67 0.8 GB 1.6 GB Runs comfortably 128K 141
Q5_K_S 5.52 0.8 GB 1.5 GB Runs comfortably 128K 144
Q4_K_M 4.85 0.7 GB 1.5 GB Runs comfortably 128K 157
Q4_K_S 4.58 0.7 GB 1.4 GB Runs comfortably 128K 163
Q4_0 4.55 0.7 GB 1.4 GB Runs comfortably 128K 163
IQ4_XS 4.25 0.7 GB 1.4 GB Runs comfortably 128K 170
Q3_K_M 3.91 0.6 GB 1.3 GB Runs comfortably 128K 179
IQ3_M 3.70 0.6 GB 1.3 GB Runs comfortably 128K 185
IQ3_XXS 3.06 0.5 GB 1.3 GB Runs comfortably 128K 206
Q2_K 2.63 0.5 GB 1.2 GB Runs comfortably 128K 222
IQ2_XXS 2.06 0.4 GB 1.1 GB Runs comfortably 128K 249
IQ1_M 1.75 0.4 GB 1.1 GB Runs comfortably 128K 266

Also worth checking

Llama 3.2 1B 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