New today本地部署POST
DeepSeek Flash 1017 hits 410 tok/s unquantized on dual RTX Pro 6000 Max-Q
A user runs unquantized DeepSeek Flash 1017 on two RTX Pro 6000 Max-Q cards and reaches 410 tok/s on coding, arguing the future of local inference is hardware-specific software tricks.
rwtenjoyer
@rwtenjoyerArchived 2 hours agoOriginally posted 2026.10.30
rwtenjoyer @rwtenjoyerOriginally posted 2026.10.30
We are now at 410 tok/s (coding) for Deepseek Flash 1017 (unquantized) on 2x RTX Pro 6000 Max-Qs Amazing how much is left on the table for each unique setup The future of local is hardware specific inference software tricks
x.com/rwtenjoyer/status/2116006915163726289Archived 2026.10.30 12:05