Builds / 4x Nvidia GB10 (unified)
4x DGX Spark (GB10, 2,200 MHz cap): DeepSeek V4.1 Flash 1017 at 116.0 tok/s code / 4,924 prefill (SGLang TP4, DSpark)
Last updated:
- GPU(s)
- 4x Nvidia GB10 (unified)
- VRAM / unified memory (GB)
- 512
- CPU
- NVIDIA GB10 (20-core Arm) ×4
Amazon · eBay (used) · Newegg - Motherboard
- 4× NVIDIA DGX Spark (ConnectX-7 200Gb/s)
Amazon · eBay (used) · Newegg - RAM
- 128 GB unified per box
- PSU
- not stated
- Other
- Four DGX Sparks (GIGABYTE AI TOP ATOM and ASUS GX10, which the author says behave the same here) joined through a MikroTik CRS812 switch; NVIDIA DGX OS; SGLang DeepSeek V4.1 branch (f80c91a4b) tensor parallel 4 under a published recipe version 2.4 (403a5aa), RDMA on both halves of each ConnectX-7 port; graphics chips power-capped at 2,200 MHz; timed from another machine on the same network
- Reported total cost
- not stated
- Power
- each GB10 capped at 2,200 MHz
Measured speed (generation tokens/sec)
| Model | Quant | Engine | tok/s | Prompt processing (tok/s) | Details |
|---|---|---|---|---|---|
| DeepSeek V4.1 Flash | DeepSeek's own release (revision 1bae0e1, 510 GB on disk): experts 4-bit MXFP4, the rest 8-bit FP8 | SGLang DeepSeek V4.1 branch (f80c91a4b) TP4, DSpark n=5, 1,048,576-token window | 116.0 | 4,924 @ 15,000-token prompt est. TTFT @8k ≈ 1.7 s (calculated from measured prefill) | ⚡ spec-decode (DSpark)context: code prompt, thinking off; 1,048,576-token window configured📏 measuredlocally test: one user, thinking off, same code prompt on every machine; middle of three runs (idle machine, 8 quiet seconds before each run); graphics chips capped at 2,200 MHz (about 2,400 uncapped) to cut power use — author's August check: writing within 1%, reading 2–4% slower; middle of three code runs 112.6–120.2 tok/s; prose 63.0 (58.6–63.3); first word 0.24 s; 16 users at once: 200 tok/s in total (about 13.8 each) against 64.0 for one user in the multi-user table (prose-style answers) |
Prompt processing = prefill speed (how fast the prompt is read) as reported by the source, with the prompt length it was measured at; blank = the source does not report it. est. TTFT @8k is calculated from the measured prefill speed (8,192 ÷ prefill tok/s), shown only when prefill was measured on a prompt of 512+ tokens. It is not a measurement and ignores load time and caching. measured TTFT is the source's own time-to-first-token. Why prefill matters for coding assistants.
measuredlocally.com hardware leaderboard (Alex, UK), DeepSeek V4.1 Flash 1017 row for four DGX Sparks, tested 20 Oct 2026. Power-capped; the model's own draft layers (DSpark) guess five tokens ahead. Native mixed MXFP4/FP8 checkpoint, not a further compression. Companion to the same author's one-Spark and two-Spark Qwen3.8 27B rows.
Blank fields = not stated in the source. We never estimate build data; GPU street prices above are separately sourced and dated.
Copy into a README or blog. Links back to this build. See badge docs (CC BY 4.0 attribution). The SVG does not invent speeds — numbers stay on the build page.
Markdown
[](https://getrigbase.com/builds/4x-dgx-spark-deepseek-v41-flash-sglang-measuredlocally/)
HTML
<a href="https://getrigbase.com/builds/4x-dgx-spark-deepseek-v41-flash-sglang-measuredlocally/" rel="noopener"><img src="https://getrigbase.com/badge/4x-dgx-spark-deepseek-v41-flash-sglang-measuredlocally.svg" alt="Measured on RigBase" height="20"/></a>
Preview: 
· How to use
Quick answers: Can 4x DGX Spark (GB10) run DeepSeek V4.1 Flash?
Speculative-decoding rows on this page are not like-for-like with plain decode. See how much MTP and DFlash really speed things up (measured pairs).