GLM-5.3-Flash gets DFlash2, a block-diffusion draft model for lossless speculative decoding

TheZachMueller · x · 2026-08-28

Inco AI released GLM-5.3-Flash-DFlash2 on Hugging Face, a speculative decoding draft model for GLM-5.3-Flash. It is not a standalone LLM — it runs inside a speculative decoding server and drafts tokens for the target model to verify.

How it works: DFlash 2 is a block-diffusion drafter that predicts a whole block of tokens in one pass, keeps top candidates at every position, and a lightweight selector traces one coherent path. Two-tap dynamic convolutions in the backbone keep drafts from decaying toward the end of the block. Decoding is lossless: greedy output matches the target exactly and sampling preserves its distribution.

Serving uses SGLang with --speculative-algorithm DFLASH; evaluation ran on 4x GB300 GPUs (TP4) with FP8 KV cache, speculation block size 8 (7 draft tokens per verification step), temperature 1.0 / top-p 0.95.

Related event: Inco AI Releases DFlash 2 Draft Model, Speeding Up GLM-5.3-Flash by 2.8x(3 posts)→

Original post →

More from Infra

Infra channel →