RTX PRO 6000 Blackwell Qwen3.5-122B Q4 token-gen (Hardware Corner)

THIRD-PARTY MEASURED (Hardware Corner llama.cpp llama-bench). Q4 GGUF, CUDA, batch-1 decode @4K; 65.6 t/s @256K on the same card. 122B dense in 96GB is capacity-bound, not speed-optimised.

Largest dense model class that fits a single consumer-grade card at usable speed; Mistral-123B Q4 does 19 t/s @4K on the same part.