Qwen3.6-27B (dense) · DDTree (Diffusion Draft Tree) · BLOCKED (hybrid GDN cache)
status: blocked
Configuration
| Model | Qwen/Qwen3.6-27B |
|---|---|
| Company | Alibaba |
| Family | Qwen |
| Parameters | 27B dense (hybrid GatedDeltaNet + full-attn) |
| Engine | DDTree research harness (github.com/liranringel/ddtree, PyTorch + transformers, batch-1) + DDTree — tree draft over the z-lab/Qwen3.6-27B-DFlash block-diffusion drafter (speculative decoding) |
| Quant / precision | BF16 (harness loads the unquantized target via AutoModelForCausalLM) |
| Why this quant | EXPERIMENTS.md P0 |
| Download | Qwen/Qwen3.6-27B |
| Context window | 4096 |
| Input modalities | text (served text-only here) |
Measured results
| Prefill tok/s | — |
|---|---|
| Decode tok/s | — |
| Peak memory (GB) | — |
| Completed | 2026-07-01 |
Full run command
scripts/bench-ddtree.sh Qwen/Qwen3.6-27B z-lab/Qwen3.6-27B-DFlash mt-bench 20 512 64,256 0.0 q27b
# Same GatedDeltaNet rollback wall as the 35B-A3B.
Notes
BLOCKED — same hybrid GatedDeltaNet wall as qwen3-6-35b-a3b-ddtree-blocked.
Qwen3.6-27B is model_type: qwen3_5 (Qwen3_5ForConditionalGeneration) — the dense member of the same
Qwen3.5/3.6 family, with the same interleaved GatedDeltaNet linear-attention + full-attention stack. The
DDTree/DFlash harness verifies speculative blocks by past_key_values.crop(start) after each block, which a
recurrent linear-attention state cannot support (nothing to slice back to an arbitrary accepted position), so
the target forward fails identically. No serving engine implements DDTree, and the harness has no
hybrid-cache/state-checkpoint path.
The measurable DDTree datapoint on the Spark is on the harness-supported non-hybrid target — see
qwen3-coder-30b-a3b-ddtree.