RTX PRO 6000 Blackwell gpt-oss-120B Q8 token-gen @4K (Hardware Corner)
221.0 tok/s
Fields
id
benchmarks/comm-rtx-pro-6000-gpt-oss-120b-q8-tg-4kMetric
decode_tok_s
Value
221
Unit
tok/s
Model
gpt-oss-120B
Methodology
THIRD-PARTY MEASURED (Hardware Corner llama.cpp llama-bench). Q8 GGUF, CUDA, batch-1 decode @4K; 123 t/s @128K on the same card. 96GB GDDR7, ~1,790 GB/s, no FlashAttention-4.
Measured by
third_party
Reproducible
yes
notes
The only single card that runs 120B+ class models locally at usable speed. $/tok-s is the worst in the consumer table for small models but unmatched capacity; ~$8,500-8,800 Oct-2026.
Notes
The only single card that runs 120B+ class models locally at usable speed. $/tok-s is the worst in the consumer table for small models but unmatched capacity; ~$8,500-8,800 Oct-2026.