Inco AI’s DFlash2 drafter boosts Qwen3.8-Flash-Next speed by 17%
Inco AI released DFlash2, an updated speculative decoding drafter for Qwen3.8-Flash-Next, offering a 17% average speedup over the original Flash-Next model. The drafter proposes token blocks in parallel and uses a sliding-window attention mechanism (2048 tokens) to reduce compute overhead. Benchmarks on an Apple M5 Max show DFlash2 achieves 75.8 tokens/second at its default settings, compared to…
Key points
- DFlash2 drafter improves Qwen3.8-Flash-Next speed by **17%** on average, measured at **75.8 tokens/sec** vs **74.5 t/s** for Flash-Next
- GGUF file size drops to **287 MiB** (Q4\(0\)) from **6x larger** MTP head, with no runtime cost increase
- Optimized for list-heavy workloads: deeper drafts (+48% speedup) hurt prose quality slightly but help structured outputs
PSA: Upgrade Flash Next to DFlash2 if you haven't! +17% on average
huggingface.co · 30 October 2026
Loading the full article…
This text was published by huggingface.co. It is reproduced here with attribution so you can read it in full; the rights remain with the publisher. Read it at the source ↗
Coverage and discussion
1source- Reddit discussionreddit.com
The headline, key points and digest above were generated by Digest AI's editorial model from the linked sources. Automated summaries can contain errors: the sources are the record. Spotted a mistake? Tell us. Published by Martin K., who runs Digest AI and handles corrections.
More in Research
All →- IDPForge presents deep learning method for intrinsically disordered proteins · 1 src
- SETI Institute and Breakthrough Listen add agentic AI to Allen Telescope Array search · 1 src
- DeepMind researcher argues AI math breakthroughs demand new progress model · 1 src
- Open-source tool REA lets AI reverse engineer software for anyone · 2 src
- OpenAI CEO Sam Altman says math onslaught won't stop despite researcher frustration · 2 src
Comments
via GitHub Discussions