deepseek-v4.1POST
Local deployment: DeepSeek 4.1 Flash 1017 hits 14 tok/s at Q2 on R9700 Pro and 9070XT
Using lucebox RAM pinning, expert tuning and DSpark Q2 quantization, the model reaches up to 14 tok/s on an R9700 Pro, 9070XT, 64GB RAM and 7700X, with only a 32k context window so far.
James Roy Reese III 🏵️✝️ 🇺🇲
@westvirginiabldArchived 12 days agoOriginally posted 2026.10.18
James Roy Reese III 🏵️✝️ 🇺🇲 @westvirginiabldOriginally posted 2026.10.18
With RAM pinning modifications of lucebox, expert tuning, and DSpark (q2 quant atm) - I've got **DeepSeek 4.1 Flash 1017 local at up to 14 tok/sec** On: R9700 Pro 9070XT 64GB System RAM 7700X Honestly remarkable. Currently only 32k context window.
x.com/westvirginiabld/status/2111703899316859220Archived 2026.10.18 16:34