Ryzen AI Max+ 395 gpt-oss-120b MXFP4 token-gen
33.7 tok/s decode (Vulkan AMDVLK)
Fields
id
benchmarks/comm-strix-halo-gpt-oss-120b-mxfp4-tgMetric
decode_tok_s
Value
33.7
Unit
tok/s decode (Vulkan AMDVLK)
Model
gpt-oss-120b
Methodology
lhl/strix-halo-testing llm-bench, llama.cpp Vulkan AMDVLK, MXFP4; pp512 386.8 (Framework Desktop-class)
Measured by
third_party
Thermal state
unknown
notes
MoE sweet spot - 117B total / 5.1B active. Range 34-50 t/s on newer tuned ROCm builds. gpt-oss-20b 47.1; GLM-4.5-Air UD-Q4_K_XL 23.4; Qwen3-235B-A22B UD-Q3_K_XL 15.1; Llama-4-Scout 19.4.
Notes
MoE sweet spot - 117B total / 5.1B active. Range 34-50 t/s on newer tuned ROCm builds. gpt-oss-20b 47.1; GLM-4.5-Air UD-Q4_K_XL 23.4; Qwen3-235B-A22B UD-Q3_K_XL 15.1; Llama-4-Scout 19.4.