gpt-oss-120b-tg128-r20.csv, max-performance campaign
By hogeheer499-commits · GitHub
- Source published
- May 7, 2026
- Test performed
- May 7, 2026
- Added to MiniPC.guru
- Sep 17, 2026
- Source reviewed
- Sep 17, 2026
- Tested hardware
- AMD Ryzen AI Max+ 395 with Radeon 8060S, 16 cores / 32 threads · Radeon 8060S Graphics, gfx1151, RADV STRIX_HALO · 128 GB LPDDR5X-8000 unified memory, about 124 GiB visible to the operating system, 512 MB assigned to the iGPU in firmware · not applicable: token generation does not read from disk. The source documents the memory but not the SSD capacity; Beelink sells the GTR9 Pro in a single 2 TB configuration
- Software / version / OS
- llama-bench, Vulkan backend · llama.cpp build 9049, commit 2496f9c14 · Ubuntu 24.04, kernel 6.19.4-061904-generic, Mesa 26.0.6
- Power profile
- tuned accelerator-performance; IOMMU disabled; firmware UMA set to 512 MB
- Test conditions
- gpt-oss 120B MXFP4 MoE, 116.8 billion parameters, 63.4 GB on disk. Quantisation MXFP4. n_gen 128, n_prompt 0, n_depth 0, so pure generation of 128 tokens at zero context depth; n_batch 2048, n_ubatch 512, n_threads 16, type_k and type_v f16, n_gpu_layers 999, flash_attn on, use_mmap off, AMD_VULKAN_ICD=RADV. Twenty repetitions, standard deviation 1.41 tok/s
Editorial review checks the source and documentation; it does not certify the measurement or mean we tested the device ourselves.






