Ryzen AI Max+ 395 Strix Halo Qwen3-235B-A22B UD-Q3_K_XL decode
15.1 tok/s (single-stream)
Fields
id
benchmarks/comm-strix-halo-qwen3-235b-a22b-udq3-decodeMetric
decode_tok_s
Value
15.1
Unit
tok/s (single-stream)
Model
Qwen3-235B-A22B UD-Q3_K_XL (235B, ~22B active)
Methodology
lhl/strix-halo-testing llm-bench; ~100 GB resident, tight but fits 128GB
Measured by
third_party
Thermal state
unknown