Mia(@MiaAI_lab):Run DeepSeek v4.1 Flash on your 3x DGX Sparks ✨ Conservative settings by default for stability: - 500k context, 750k KV cache - prefill is ~1600-1700 tok/s in all ctx ~38 tok/s prose in single stream ~79 tok/s prose in 4 concurrent streams Full prose decode tok/s numbers in below post. Get it here:

Mia

@MiaAI_lab

Building with AI & LLMs | Insights, recipes, tools & honest experiments.

参加 July 2022

413 フォロー中 36.1K ファン

Mia@MiaAI_lab

2026.09.11 08:21

Run DeepSeek v4.1 Flash on your 3x DGX Sparks ✨ Conservative settings by default for stability: - 500k context, 750k KV cache - prefill is ~1600-1700 tok/s in all ctx ~38 tok/s prose in single stream ~79 tok/s prose in 4 concurrent streams Full prose decode tok/s numbers in below post. Get it here: