Inco AI releases DFlash 2 draft model for Qwen3.8-Flash-Next
DFlash 2 draft model is released as a repository that runs inside a speculative decoding server and drafts tokens for the target model Qwen3.8-Flash-Next. It predicts a whole block of tokens in a single pass, keeps the top candidates at every position, and uses a lightweight selector to trace one coherent path. Two‑tap dynamic convolutions in the backbone keep the draft from decaying toward the…
Key points
- dflash 2 drafts up to 3 tokens per verification step with block size 4
- sampling uses temperature 1.0, top‑p 0.95, top‑k 20 and max 4096 new tokens
- benchmarks run 128 samples at concurrency 1 and 1,024 at concurrency 8 and 32 on an nvidia gb300
The target is the official FP8 checkpoint Qwen/Qwen3.8-Flash-Next-FP8. Runtime uses SGLang on one NVIDIA GB300 (TP1). Speculation block size is 4, giving three draft tokens per verification step, limited by the current QSA implementation. Sampling follows the officially recommended parameters: temperature 1.0, top‑p 0.95, top‑k 20, with a maximum of 4096 new tokens. Benchmarks include 128 samples at concurrency 1 and 1,024 samples at concurrency 8 and 32, comparing autoregressive decoding, Flash‑Next’s built‑in MTP, and DFlash 2, with token‑per‑second speedups reported.
incoai/Qwen3.8-Flash-Next-DFlash2
huggingface.co · 19 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 Generative AI & Models
All →- OpenAI briefly listed three aurora-alpha models in its API, later removed · 3 src
- DeepSeek releases V4.1-Flash-1017 update with lower cost and higher DeepSWE score · 23 src
- Researchers release Pearl-Chat, a Luganda language model with 8.5M parameters · 1 src
- Researchers swap Gemma 3 270M embedding table for 16 MB frozen vocabulary · 1 src
- Mistral AI releases Timestral forecasting model family, beats Google TimesFM-3 · 1 src
Comments
via GitHub Discussions