Builds / 2x Nvidia GB10 (unified, DGX Spark)
2x DGX Spark (AI TOP ATOM + ThinkStation PGX): DeepSeek V4.1 Flash 1017 EXL3 2.9 bpw on Mia's TensorFold recipe, independently measured at 66.1 tok/s, prefill 2,135 tok/s at 32K
Last updated:
- GPU(s)
- 2x Nvidia GB10 (unified, DGX Spark)
- VRAM / unified memory (GB)
- 256
- CPU
- Nvidia GB10 Grace (Arm) per node
Amazon · eBay (used) · Newegg - Motherboard
- Gigabyte AI TOP ATOM + Lenovo ThinkStation PGX
Amazon · eBay (used) · Newegg - RAM
- 128 GB LPDDR5X unified per box (256 GB total)
- PSU
- not stated
- Other
- 2 GB10 boxes (1x Gigabyte AI TOP ATOM + 1x Lenovo ThinkStation PGX) linked by CX7 RDMA (~109 Gb/s), driver 580.178.04, kernel 7.0.0-1019-nvidia, GPU clock capped at 2,200 MHz; Mia's AI Lab recipe commit c7f2b85 (9 Nov 2026), TensorFold native Zig engine (BertholomusAI deepseek-v41-zig fbbc971) + the recipe's engine patches, TP=2, DSpark drafts, 4 streams, context 524,288
- Reported total cost
- not stated
- Power
- not stated
Measured speed (generation tokens/sec)
| Model | Quant | Engine | tok/s | Prompt processing (tok/s) | Details |
|---|---|---|---|---|---|
| DeepSeek V4.1 Flash | 1017 checkpoint, Mia EXL3 EAQ v1 2.9 bpw (rev 45ee5d0f) + DSpark drafts | TensorFold Zig, 2 nodes TP=2, 1 request | 66.1 | 2,135 @ 32,768-token prompt est. TTFT @8k ≈ 3.8 s (calculated from measured prefill) measured TTFT: 4.1 s @ 8,192-token prompt | ⚡ spec-decode (DSpark)context: 524,288 ctx, 512 generated tokens, thinking off📏 Mean of code (66.5) and prose (65.7) medians, 2 runs each after 1 warm-up, temperature 1.0, seed 42, no prompt cache; draft acceptance 57%; 58.4-68.9 tok/s with 16K-128K prompts; measured 2026-11-09 |
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.
Independent re-run of Mia's AI Lab 2-Spark recipe by niklaslenz.com (Results v2), measured 9-10 Nov 2026, with all raw files published. Uses its own prompts and sampling, so the author says it is not comparable with Mia's README figures (RigBase's Mia-recipe build lists 74.9). Several requests at once (total): 1 = 55.8, 2 = 75.5, 4 = 99.3, 8 = 97.6 tok/s. Peak memory rise 108.7 GiB per box with a 128K prompt. Own quality runs (thinking off): MMLU-Pro 66.8%, GPQA Diamond 38.9%, IFEval 88.5%; GPQA 30 with thinking 83.3%.
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/2x-dgx-spark-deepseek-v41-flash-1017-mia-recipe-independent-niklaslenz/)
HTML
<a href="https://getrigbase.com/builds/2x-dgx-spark-deepseek-v41-flash-1017-mia-recipe-independent-niklaslenz/" rel="noopener"><img src="https://getrigbase.com/badge/2x-dgx-spark-deepseek-v41-flash-1017-mia-recipe-independent-niklaslenz.svg" alt="Measured on RigBase" height="20"/></a>
Preview: 
· How to use
Quick answers: Can 2x 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).