Qwen3.6-35B-A3B on a single DGX Spark: 25.5 → 45.2 tok/s

Published 25 Oct 2026 · Updated 28 Oct 2026 · measured 12 Jul 2026

1× GB10 (DGX Spark) · 16-bit · median of 5 at temperature 0 · vLLM development build · protocol stated below

The most useful benchmark for most Spark owners is not the biggest model — it is what a strong mid-size MoE does on one box, and how far a plain eager-mode config is from what the box can do. Answer: a long way. Same silicon, same weights, config changes only:

That is +77% on code from configuration alone — no new hardware, no different weights. The winning settings differ from model to model and have to be measured, not copied.

Protocol (shorter than our current n ≥ 15 standard): one prompt per workload (dialogue, code), 3 warm-up requests then 5 timed, median reported; 256 output tokens; temperature 0; one server launch per row; rate = tokens ÷ total request time. No confidence intervals and no repeat launches. Within each set of five the spread was small (code 43.2–46.7 in the last row). Measured 12 Jul 2026 on a vLLM development build. The model is Hugging Face Qwen/Qwen3.6-35B-A3B; an earlier version of this page called it “Qwen3.6-Next”, which is not a real model name.

The multi-node question

This entry is deliberately single-node. The same weights were also measured on a different engine, SGLang, in a separate test: one request at a time, 17.45 tok/s on one node and 39.78 tok/s across eight (tensor plus expert parallel), 2.3× faster for one user. Those runs used SGLang's own benchmark (a 128-token prompt and a 256-token answer, one request), not this page's protocol, so compare them with each other, not with the vLLM numbers above. With only ~3B parameters active per token you might expect little gain from more nodes, since each node already reads little data per token; on SGLang it still gained. Clustering pays most for models too big for one box, and for aggregate throughput. In our September hand-tuned tests of GLM-5.3-Flash (dense attention, no speculation), 8 nodes beat 4 by 34% single-stream once CUDA graphs removed launch overhead, though the vendor's documented 4-node configuration (with its built-in speculative decoding) later beat both (39.13 vs 35.77 tok/s). Measure before buying a second box for speed.

What the gap means

An eager-mode launch runs the 25.5 line. The 45.2 line is the same computer. The difference is knowing what to configure and how to verify the win survives restarts — two of the standard optimizations can silently cancel each other.

Free to read and share. Numbers are from hardware we've measured directly; your engine versions and configs will differ — measure your own, and count only delivered tokens.