Benchmarks · Measured, not quoted
Qwen3.8 27B on DGX Spark
These are our own measurements on our production hardware — the same machines serving api.axforge.ai. Every number carries its configuration and date; nothing here is a vendor quote.
Results
What we measured
| Setup | Single-stream decode |
|---|---|
| Qwen3.8 27B (NVFP4) with multi-token-prediction speculative decoding | 15.9 tokens/s |
| Qwen3.8 27B (NVFP4), speculative decoding off | 5.8 tokens/s |
| Qwen3.6 35B A3B (MoE, ~3B active) — for comparison | 30.7 tokens/s |
- Measured on our production DGX Spark node, single-stream, 2026-08.
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 | 65,536-token context, up to 4 concurrent sequences, vLLM-based serving with multi-token-prediction speculative decoding. |
| Measurement | Single-stream decode throughput — tokens generated per second for one request. |
Limitations
What we have not measured
We publish only what we measured: no time-to-first-token numbers, no concurrency curves, no third-party quality benchmarks — for model quality, see the model's official card. When we measure more, it lands here with the same labelling.
What it means
Reading the numbers
The MoE comparison shows the trade honestly: Qwen3.6 35B A3B decodes faster (fewer active parameters per token), Qwen3.8 27B is the stronger daily driver. Speculative decoding is why the dense model is usable single-stream — 2.7× over the baseline on the same silicon.
The same machine is rentable: DGX Spark, €495/month — or just call the model on the shared API from €0.29 / 1M input tokens.
Related