Qwen3.8-27B DSpark
Throughput
Decode tok/s after a fixed context fill (PBM) — same agent-style workload as bench v2.
Published evals
Vendor-card figures for qwen/qwen3.8-27b — not this pack. Not measured on this Spark.
- Terminal-Bench 2.1
- 73.0% Qwen model card ↗ 2026-08-14 Terminal-Bench 2.1 with Terminus (vendor).
- SWE-bench Pro
- 61.7% Qwen model card ↗ 2026-08-14 Claude Code harness; refined SWE-bench Pro set (vendor).
- LiveCodeBench
- 90.3% Qwen model card ↗ 2026-08-14 LiveCodeBench v6 (vendor).
- GPQA Diamond
- 89.2% Qwen model card ↗ 2026-08-14
Independent board on Artificial Analysis ↗ — we do not copy their scores.
Recipe
- Profile
- radixark-qwen3-8-27b-dspark-eugr
- Engine
- vLLM
- Context
- 64k · fp8 KV
- Served as
- qwen3.8-27b-dspark
- Draft
- DSpark · n7
Run on your Spark
single node
shell
spark inference up radixark-qwen3-8-27b-dspark-eugr
Why we run it
Public identity for NVFP4 + DSpark (Doopeworld drafter). Weights stay at radixark/qwen3.8-27b so MTP and DSpark do not collapse to one site row.
Bench notes
PBM 4k @ 22.0 / 50k @ 10.1 / 100k @ 6.4 tok/s — perfbench-metrics — profile=radixark-qwen3-8-27b-dspark-eugr