DeepSeek V4.1 Flash on 4× CMP 170HX: 296 tok/s single user

Open recipe runs DeepSeek V4.1 Flash on four CMP 170HX cards: 296 tok/s structured for one user, 796 tok/s aggregate for eight, 1M context, with 189 GiB Engram tables streamed from NVMe.

gnmac

@GNmacArchived 10 days agoOriginally posted 2026.10.21

gnmac @GNmacOriginally posted 2026.10.21

Open original on X

DeepSeek V4.1 Flash 1017 on 4× CMP 170HX — open recipe 🛠️⚡ 1 user: 296 tok/s structured / 150 prose 8 users: 796 tok/s aggregate 1M context · KV 1.57M tokens Prefill 1,759 tok/s 189 GiB Engram tables streamed off NVMe 180 W per card https://github.com/GNmac-lab/deepseek-v41-flash-1017-cmp170hx

x.com/gnmac/status/2112903602457165971Archived 2026.10.22 12:01

Open original on X