Qwen3.6-27B NVFP4 + DFlash: my daily driver (until 3.8)
Before Qwen3.8 landed, Qwen3.6-27B (NVFP4) with DFlash speculative decoding was my favourite local model on the DGX Spark (GB10) — the dense 27B quality I wanted, finally at a speed I could live with. The earlier FP8 build of this same model only managed ~8–16 tok/s; NVFP4 + DFlash changed the picture completely.
The headline: ~6× over the FP8 baseline
Same workload (20 requests, concurrency 2, ~512 in / 128 out), FP8 baseline vs NVFP4 + DFlash:
| Metric | FP8 baseline | NVFP4 + DFlash | Change |
|---|---|---|---|
| Output throughput | 16.74 tok/s | 104.56 tok/s | 6.25× faster |
| Request throughput | 0.131 req/s | 0.817 req/s | 6.24× faster |
| Mean latency | 15.29 s | 2.45 s | 84% lower |
| p95 latency | 15.37 s | 2.45 s | 84% lower |
| TTFT | 589 ms | 1,020 ms | 73% slower |
Roughly a 6× jump in throughput and an 84% drop in end-to-end latency. The one regression is TTFT: the speculative draft plus a busier scheduler add ~430 ms before the first token. In interactive chat that’s not noticeable, and total latency still collapses because decode is so much faster.
A second run, taken as the median of three, landed a touch lower but told the same story — 16.04 → 83.04 tok/s, a 5.2× jump, latency p50 down 81%. Across runs the stack sat in a ~83–105 tok/s band, every one of them 5–6× the FP8 baseline.
Measurement notes
A win this size needs a trustworthy yardstick, so the method matters:
- Report the median of three identical runs, not a single shot. The FP8 baseline alone drifted ~10% run to run (15.16–16.79 tok/s), so any single number lies a little.
- Treat changes under ~5–10% as noise, not improvement — that’s the run-to-run spread on this box. A 6× jump clears that bar with room to spare; a firmware tweak claiming +8% does not.
- Save the winning config as the new baseline and only re-benchmark after a deliberate change (firmware, driver, recipe). Otherwise you’re chasing noise.
This was the model that made local dense-27B reasoning genuinely usable for me — right up until Qwen3.8-27B and DFlash2 raised the bar again.