D3D
D3D Cuts LLM Inference Time by Up to 7.48x, but Must Escape the Paper
Fudan University's D3D parallelizes draft generation and target-model verification, delivering average Qwen3 speedups of about 5x and a peak of 7.48x.
Back to home
speculative decoding
6 articles tagged with this topic
Qwen
Qwen's SpecOPD Combines Branch-Parallel Training with Speculative Decoding
Qwen's SpecOPD accelerates inference by combining speculative decoding with branch-parallel training, achieving gains of up to 3.56x in generation spe
DFlash 2
2.26x Speedup Is Real. The '8x' Number Is Fake — An Honest DFlash 2 Test
DFlash 2 hits a real 2.26x speedup. But a developer's 3-day benchmark debunks the viral '8x' number — AI perf claims can be marketing traps.
llama.cpp
Reddit 用户让本地 AI 提速 65%，但官方还没接盘
llama.cpp fork with DSpark PC Tree speculative decoding pushes Qwen3 ~65% faster on RTX 5090. Not merged, but local AI is getting cheaper and faster.
Meta
Meta's Muse Glimmer 30B Hits 3.3x Speedup on Mac as Open Source Closes the Gap
Developer A-Rahim uses speculative decoding to push Meta's Muse Glimmer 30B up to 3.3x faster on M4 Pro, with output exactly matching the original.
Gemma 4
Gemma 4 Has Hidden MTP Heads Disabled by Google at Launch
A developer found multi-token prediction weights inside Gemma 4's LiteRT files; Google confirmed MTP exists but was intentionally disabled.
Apr 73 min readopcnew Editorialreddit.com