DFlash 2 pushes parallel drafting further: over 20% more output per verification pass for ~1% added latency

DFlash 2: Keep Drafting Parallel

Inco AI released DFlash 2, adding a lightweight path selector on top of parallel speculative decoding. Instead of keeping only the top-1 candidate per position, it picks a coherent path from the top 16, raising accepted tokens per verification from 4.27 to 6.79. On Qwen3.8-27B with SGLang, throughput reaches 2.7–3.4× autoregressive decoding at batch size 1, with roughly 1% added cycle latency. SGLang, vLLM, llama.cpp, and oMLX already support it; DFlash models have been downloaded over 3.5 million times on Hugging Face.

Why it matters: DFlash 2 is a clear technical improvement on an already-adopted inference method, with measured results. Score isn't higher because this is a single technical blog post, not a model launch or product release — its reach is limited to the inference-stack crowd.