DeepSeek-V4.1-Flash 552B MoE on 4x DGX Spark TP4 - 77.2 tok/s c1 on PEAK 52 code · 72 tok/s on a warm code run · 47 math · 39 reasoning · 23 prose

Accelerated Computing DGX Spark / GB10 User Forum DGX Spark / GB10 Projects
1

Deployment checkpoint. Got it running on 4 x DGX Sparks working on tuning now for speeds this REPO will be changing throughout the day feel free to join me for PRs and Issues.

🔥 DeepSeek-V4.1-Flash 552B MoE (769B w/ Engram) · MXFP4 · 4x DGX Spark TP4 · vLLM + DSpark + CUDA graphs

Re-benched warmed + streaming (fixed 8-category prompt set, C1 to C6):
⚡️ 77.2 tok/s PEAK single-stream on counting · 71 on structured tables (DSpark runs hot: up to 5.95 of 6 tokens accepted per step)
💻 52 code in the bench · 72 tok/s on a warm code run · 47 math · 39 reasoning · 23 prose
🚀 214 tok/s aggregate at 6 streams · 143 on code
🧠 1.03M-token KV · 300K ctx · 1M ctx proven (1.08M-token pool) · 0.27s TTFT
👀 Vision + 🛠️ tool calling on
💾 203 GB of Engram tables stay on disk, so 552B fits in 4x 128 GB unified memory
🔧 5 tok/s eager ➜ 41 with CUDA graphs + DSpark ➜ 77 after un-latching 2 GPUs stuck below 1 GHz (unplug the Spark to clear it, a reboot won’t)

Day-0. 552B-class reasoning on desk hardware 🖥️

2

Will you be working on a dual spark version too?

3

Great stuff, I did not expect this to be working for a week tbh.

Do you have any pre-fill numbers as your repo only show decode speed?

4

I will get those added now an will post in here give me a few mins

5

amazing! i actually got 2 more sparks to a total of 4. did you have to run a Router to make this work? or just daisy chaining the connectX7 cables? ive been trying to score the mikrotik crs804 router but its like impossible to find nowadays

6

Got it running TP3 now too ! Same repo