Benchmarks
Qwen3.8 27B on DGX Spark
Labelled community benchmark data for the hardware class, plus the exact serving configuration behind the API. Our own tuned measurements are published here once a clean run is complete. The early production numbers were captured on a busy, untuned node and undersold the machine, so they are not shown.
Community data
What the hardware class does
| Setup | Single-stream decode |
|---|---|
| Qwen3.8-27B (NVFP4 + MTP), SparkBench PBM @ 4k, single GB10 | 25.1 tokens/s |
| Qwen3.6-35B-A3B (NVFP4 + MTP, MoE ~3B active), SparkBench PBM @ 4k | 86.3 tokens/s |
- Community measurements from SparkBench, labelled, not ours. Full tables, llama.cpp data and the MoE-vs-dense story on the GB10 page.
Methodology
Exactly how
| Hardware | NVIDIA DGX Spark: GB10 Grace-Blackwell, 128 GB unified memory. The production node, not a lab rig. |
|---|---|
| Model | Qwen3.8 27B, NVFP4 quantization, the exact build served on the API. |
| Serving config | 262,144-token context, up to 6 concurrent sequences, vLLM-based serving with multi-token-prediction speculative decoding. |
Our numbers
Where AxForge's own measurements are
Coming with a clean, tuned run. Our early measurements were captured while the node carried other load, with no tuning pass. They undersell the machine, so they are not shown. When we benchmark properly, the results land here, dated and with their full configuration.
What it means
Reading the numbers
The MoE comparison shows the trade honestly: Qwen3.6 35B A3B decodes several times faster (fewer active parameters per token), while Qwen3.8 27B is the stronger daily driver. Serving-stack choices, speculative decoding above all, move dense-model results as much as the silicon does.
The same machine is rentable: a dedicated DGX Spark with full SSH access, from €0.85/hour by the hour, week, month or year, or just call the model on the shared API from €0.32 / 1M input tokens.
Related