Open-PerfectBlend
AI overview
Sign in and the AI will write an overview from our coverage.
Headlines · 2
- D3D enables asynchronous speculative decoding with diffusion drafting
The arXiv paper introduces D3D, a standalone diffusion drafter that removes the dependency between successive drafting rounds by combining blockwise denoising with auxiliary autoregressive supervision. Its independence from target-model hidden states allows successor blocks to be drafted while verification runs, while progressive layer pruning reduces packed-drafting cost. The authors report average decoding speedups of 5.00× and 5.71× across seven benchmarks with Qwen3-8B and Qwen3-32B targets, plus 7.48× on GSM8K with Qwen3-32B; these figures remain author-reported.
Hugging Face · Papers · 🔥 0 - Tencent releases DSpine speculative decoding drafter for Qwen3-8B
Tencent released the Qwen3-8B-DSpine-b16 model card and inference code for DSpine, a parallel drafter that uses adjacent causal feature injection across network depth. The model card reports longer acceptance lengths than DSpark and DFlash2 across seven math, code, and chat benchmarks after retraining all three methods with the same configuration. The paper authors also report that DSpine provides 23.3% higher average throughput than DFlash in SGLang serving tests, although the supplied reports contain no independent replication.
腾讯混元 · Hugging Face · 🔥 0
Experience and discussion from the community
Share my Open-PerfectBlend experienceAsk about Open-PerfectBlend
Nobody has shared their experience with Open-PerfectBlend yet.