DFlash 2
- Announcement — DFlash 2: Keep Drafting Parallel, Inco AI, 18 August 2026. https://inco.ai/blog/dflash2/
- Drafter (safetensors) — https://huggingface.co/incoai/Qwen3.8-27B-DFlash2
- Drafter (GGUF) — https://huggingface.co/incoai/Qwen3.8-27B-DFlash2-GGUF
- Mirror — https://huggingface.co/z-lab/Qwen3.8-27B-DFlash2
- Reference implementation — https://github.com/z-lab/dflash
- vLLM support — PR #52816, [Spec Decode] DFlash2: local convolution + candidate selector. https://github.com/vllm-project/vllm/pull/52816 (open at time of writing)
- llama.cpp support — PR #27342, spec : add DFlash2 support (local convolution + candidate selector). https://github.com/ggml-org/llama.cpp/pull/27342 (open)
Published figures cited for comparison (Inco AI, H200, SGLang, block size 8, temperature 1.0
with xhigh reasoning): acceptance length 4.80 mean for Qwen3.8-27B against MTP 4.28 and a
community DSpark drafter 3.62; throughput 2.7-3.4x autoregressive at concurrency 1.
Models
- Target — https://huggingface.co/Qwen/Qwen3.8-27B and the
-FP8variant - NVFP4 — https://huggingface.co/unsloth/Qwen3.8-27B-NVFP4
- int4 AutoRound — https://huggingface.co/Pilcothink/Qwen3.8-27B-MixedInt4-AutoRound
- GGUF — https://huggingface.co/ggml-org/Qwen3.8-27B-GGUF
- DSpark drafter used here — https://huggingface.co/Doopeworld/Qwen3.8-27B-DSpark-vLLM (note: Inco AI benchmarked https://huggingface.co/RadixArk/Qwen3.8-27B-DSpark, a different community checkpoint — see the caveat in DFLASH2.md)
Engines
- vLLM 0.27.1 — https://github.com/vllm-project/vllm
- llama.cpp — https://github.com/ggml-org/llama.cpp
Evaluation data
- GSM8K — https://huggingface.co/datasets/openai/gsm8k (test split, n=200, seeded sample)
- MMLU — https://huggingface.co/datasets/cais/mmlu (test, n=400 across 56 subjects)
Prior art referenced by DFlash 2
- Canon Layers; Dynamic Short Convolutions; Convolution for Large Language Models — cited in the announcement as the basis for the two-tap dynamic depthwise convolution.
- Modal, Speculation Is All You Need — cited on speculative decoding for low-latency serving.