Models
Qwen3.6-27B
2.29× mean on the 8-bit target with a strong block-15 community drafter.
- Made by
- Alibaba Qwen
- Size
- 27B
- Architecture
- Hybrid: linear + full attention
- Features
- Tool callingThinking
Run it
mlx-dspark serve --model mlx-community/Qwen3.6-27B-8bit
Serves an OpenAI + Anthropic API on http://127.0.0.1:8080. For a one-off answer, use mlx-dspark generate --model mlx-community/Qwen3.6-27B-8bit --prompt "…".
Drafter
satgeze/Qwen3.6-27B-DSparkresolves automatically with--mode auto(the default).
The drafter downloads with the model the first time you run it, and loads 4-bit quantized (acceptance doesn't depend on the drafter's precision).
Reading the prompt
Prefill runs at 126 tok/s, so a 20,000-token prompt takes about 159 s the first time. After that the prefix cache skips it.
Measured 2026-07-25 on an M4 Pro, 48 GB with mlx 0.32.0. This predates the September 2026 verify kernels, which raised every 8-bit model re-measured since, so it is probably conservative.
Qwen3.6-27B with satgeze's community DSpark drafter: a block-15 head (7 is typical), trained against the bf16 target and warm-started from z-lab's DFlash head for the same model.
The rule of thumb it follows: match the target's precision to what the drafter was trained against. A bf16-trained drafter wants the 8-bit target. The 4-bit target resolves the same drafter and should trade ratio for absolute speed, but it isn't measured.
These numbers predate the September 2026 verify kernel, which moved other 8-bit targets from cap 4 to 7.
Numbers not matching your Mac? They shouldn't be identical: the draft cap is derived from your machine's measured curves. Run mlx-dspark benchmark --model mlx-community/Qwen3.6-27B-8bit --trials 3 to get your own. See how these are measured.