Qwen3.8-27B DFlash2 GGUF: Faster Speculative Decoding Setup

Qwen3.8-27B (DFlash2 GGUF)

Accelerator for Qwen3.8-27B

A diffusion drafter for speculative decoding. It works in tandem with the target Qwen3.8-27B, speeding up generation without loss of quality.

You can read about the base qwen3.8-27b in this post; in benchmarks it confidently outperforms Qwen3.6-27B, Qwen3.7-Plus, and Muse Glimmer-30B, and in many categories surpasses Opus4.6 Max.

System requirements

If you have 24 GB VRAM, take the drafter in Q4_K_M and the target model in Q4_K_M or IQ4_XS (this is optimal).

If you have 16 GB, try the drafter Q4_K_M + target Q3_K_M (context up to 16k).

If you have 48 GB or more, take the drafter Q8_0 and the target Q8_0.

BF16 is already for 128 GB+.

️DFlash2 requires about 2–3 GB on top of the drafter weights (hidden-state taps and verification buffers). On a 16 GB GPU it may be tight. The Q2_K drafter version (~700 MB) saves ~400 MB and gives +20k context without loss of quality.