The GB10 benchmark database
What big models actually run at on DGX Spark / GB10 hardware, from one node to eight — measured as tokens delivered to the client, not server-log numbers. The October entries follow the methodology (n ≥ 15, 95% confidence intervals); the July entries used shorter protocols, stated on each entry, on older engine builds. If a number here surprises you, that's the point: many published numbers don't say which tokens they count or over what window.
| model | size | nodes | quant | decode (code) | measured · protocol | detail |
|---|---|---|---|---|---|---|
| Kimi K2-class | ~1T MoE | 8 × GB10 · 100GbE | 4-bit | 27.8 tok/s | Jul 2026 · n=15, 95% CI | full entry → |
| Qwen3.6-35B-A3B | 35B MoE (3B active) | 1 × GB10 (DGX Spark) | 16-bit | 45.2 tok/s | Jul 2026 · median of 5 | full entry → |
| GLM-5.3-Flash-class | ~320B MoE | 4 × GB10 · 100GbE | 4-bit | 98.8 tok/s | Oct 2026 · n=15, 95% CI | full entry → |
| GLM-5-class | 753B MoE | 8 × GB10 · 100GbE | mixed 4-bit | 16.5 tok/s (no code/prose split; tuning unfinished) | Jul 2026 · best of 3 | full entry → |
| MiniMax M3-class | ~400B MoE + vision | 8 × GB10 · 100GbE | 4-bit | 34.9 tok/s | Jul 2026 · n=15, 95% CI | full entry → |
All entries are single-stream decode on code workloads unless labeled otherwise; each full entry breaks out prose, prefill, aggregate, and long-context behavior where measured. Numbers are from our own cluster of eight GB10 nodes (seven ASUS GX10 and one DGX Spark) on an ordinary 100 Gbps switch — commodity networking, not a lab exotic, though most owners run fewer nodes. Engine builds change fast: check the measured date before comparing with your own numbers.
Why this database exists
Few people publish reproducible multi-node Spark numbers for models this size; knapcio's open GLM-5.3-Flash setup is a notable exception, and our Flash entry now builds on it. Every entry here records a measured tuning journey: what the model did out of the box, what class of problem was in the way, and what it runs at now. The journeys are described; the exact fixes are what the diagnostic delivers for your cluster.
Want a model measured?
We add entries as we tune them. If you need a number for a specific model on Spark hardware before you buy or deploy — that's a service we sell, with the harness output included.
Ask for a benchmark →