Builds / Nvidia GB10 (unified)
DGX Spark: Nemotron 3.5 Bolt 74B-A9B on Mia's TensorFold recipe, independently measured at 67.9 tok/s (prose 72.5), prefill 2,496 tok/s at 8K
Last updated:
- GPU(s)
- Nvidia GB10 (unified)
- VRAM / unified memory (GB)
- 128
- CPU
- Nvidia GB10 Grace (Arm)
Amazon · eBay (used) · Newegg - Motherboard
- not stated
- RAM
- 128 GB LPDDR5X unified
- PSU
- not stated
- Other
- 1 DGX Spark, driver 580.178.04, kernel 7.0.0-1019-nvidia; Mia's AI Lab Nemotron-3.5-Bolt-TensorFold-DGX-Spark recipe (commit 6b064a31, 1 Nov 2026), TensorFold v1.1.6 (eea9dc9c) + Mia patch 0002 + libbolt, LiLiCorr drafts, 16 server slots, 262,144 ctx
- Reported total cost
- not stated
- Power
- not stated
Measured speed (generation tokens/sec)
| Model | Quant | Engine | tok/s | Prompt processing (tok/s) | Details |
|---|---|---|---|---|---|
| Nemotron 3.5 Bolt 74B-A9B | IQ2_XXS + NVFP4 (NVIDIA 24GB) + LiLiCorr drafts | TensorFold v1.1.6 (Mia recipe 6b064a31), 1 request | 67.9 | 2,496 @ 8,192-token prompt est. TTFT @8k ≈ 3.3 s (calculated from measured prefill) measured TTFT: 3.3 s @ 8,192-token prompt | ⚡ spec-decode (LiLiCorr)context: 262,144 ctx, 512 generated tokens, thinking off📏 Mean of code (63.4) and prose (72.5) medians, 2 runs each after 1 warm-up, temperature 1.0 / top_p 0.95, seed 42, no prompt cache; draft acceptance 35.5%, mean accepted length 3.0; 52.0-74.7 tok/s with 16K-128K prompts; measured 2026-11-02 |
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 recipe by niklaslenz.com (Results v2), speed measured 2 Nov 2026, quality tests 10 Nov, raw files published. Its prose result (72.5) lands 0.9% above Mia's README prose figure (71.8); its code prompt differs from hers, so the code numbers are not comparable. Several requests at once (total): 1 = 58.4, 2 = 82.6, 4 = 118.1, 8 = 151.8 tok/s. Compare with sxuff's llama.cpp + LiLiCorr recipe on one GB10 (80.4 greedy).
Blank fields = not stated in the source. We never estimate build data; GPU street prices above are separately sourced and dated.
Quick answers: Can DGX Spark (GB10) run Nemotron 3.5 Bolt 74B-A9B?
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).