deepseek-v4.1IMAGE

DeepSeek V4.1 Flash hits 669 tok/s decode with BRRR recipe on RTX PRO 6000 Max-Q

Using fraserprice's BRRR recipe with fused decode kernels on b12x vLLM, DeepSeek V4.1 Flash reached 669 tok/s decode at batch 1, 9ms per step, 174ms TTFT and 11K prefill.

Dataphile

@Dataphil3Archived 4 days agoOriginally posted 2026.10.26

DeepSeek V4.1 Flash hits 669 tok/s decode with BRRR recipe on RTX PRO 6000 Max-Q

Dataphile @Dataphil3Originally posted 2026.10.26

Open original on X

@deepseek_ai V4.1-Flash at 669 tok/s decode, batch 1, on @nvidia RTX PRO 6000 Max-Q cards. @fraserpricee's BRRR recipe: fused decode kernels on b12x vLLM. 9 ms per step, 174 ms TTFT, 11K prefill. Recipe: https://github.com/fraserprice/DeepSeek-4.1-Flash-1017-BRRR Run: https://www.localmaxxing.com/runs/cmvoeshfj01ybng018exqqvw5 #LocalLLM

x.com/dataphil3/status/2114488331337703852Archived 2026.10.26 12:03