deepseek-v4-flash

Episodes (3)

Trajectory notes

• 2026-08-14T18:37:13Z: deepseek-v4-twinspark-dgx-serving closed (faded) — The recipe converges with Scott’s hardware-aware local-inference practice by making accelerator placement, precision, interconnect, KV-cache policy, and speculative decoding explicit—and its reproduced 41–44 tok/s performance p
• 2026-08-13T11:25:26Z: deepseek-v4-flash-vision-adapter closed (faded) — The project concretely converges with Scott’s composable-bespoke pattern by grafting a narrow, trainable vision adapter onto a frozen commodity model, and it is directly testable on his hardware-aware local-inference substrate.