DynaDFlash: diffusion drafter boosts speculative decoding acceptance length by 5%
charles_irl · x · 2026-11-13
A team, backed by $5,000 in compute credits from Modal, built DynaDFlash on top of DFlash, exploring diffusion language models for speculative decoding.
The key idea: they trained a diffusion drafter that dynamically decides which context from the target model should be injected at each layer to draft tokens most effectively.
On the MetaMathQA benchmark, the model achieves a 5% higher acceptance length and 4.1% more throughput over the baseline. The full blog post, source code, and model weights are all publicly available.
More from Infra
- After Microsoft outage, open-weight model torrent seeders hit record 30,000 — WolframRvnwlf · 2026-11-13
- AWS disputes outage reports: services 'have been fine', check Health Dashboard — Kyrannio · 2026-11-13
- The math says GPU export controls are working: only a sliver of Nvidia output leaks — bookwormengr · 2026-11-13
- NIDO ships fully offline open-source Android AI assistant with encrypted local RAG and E2EE mesh — NIDO_007 · 2026-11-13
- NVIDIA researcher mocks US compute roadshow: 'Huawei and DeepSeek just do their own thing' — suchenzang · 2026-11-13
- Azure SQL Always Encrypted: running range queries on data the database can't read — adnan_hashmi · 2026-11-13