DSpark: Confidence-Scheduled Speculative Decoding with Semi-Autoregressive Generation

DSpark was submitted to arXiv on 6 July 2026 by a 33-author team from DeepSeek-AI, and it targets a specific failure mode of parallel drafting in speculative decoding: when a draft model proposes several tokens simultaneously it cannot condition each token on the one before it, so the acceptance rate decays sharply as the proposal grows longer. The paper's answer has two parts. The draft model is semi-autoregressive — a parallel backbone coupled with a lightweight sequential module that restores token-to-token dependency without paying the full cost of autoregressive drafting. On top of that, verification length is not fixed: a confidence-scheduled verifier estimates the survival probability of each proposed continuation and adjusts how many tokens it verifies, factoring in current system throughput, so the batch does not burn compute verifying drafts unlikely to survive. The headline result is a 60 to 85 percent increase in per-user generation speed once DSpark is integrated into the DeepSeek-V4 serving system under real traffic, rather than on a synthetic single-request benchmark — which is the number most speculative-decoding papers do not report. The work ships as considerably more than a paper. The accompanying DeepSpec repository, at roughly 7,100 stars and 666 forks, contains data preparation utilities, draft-model implementations for three algorithms (DSpark, DFlash and Eagle3), the full training pipeline with configuration files, evaluation scripts, and released checkpoints on Hugging Face at 4B, 8B and 14B scales against Qwen3 and Gemma target models. Evaluation spans GSM8K, MATH-500, AIME25, HumanEval, MBPP, LiveCodeBench, MT-Bench, Alpaca and Arena-Hard-v2, which makes the repository usable as a comparison harness for draft-model designs rather than only a description of one.

Topic

ML

Level

Advanced

Format

Paper

Cost

Free

Duration

~60-90 minutes to read; days if you run the released DeepSpec training and evaluation pipeline

Provider

DeepSeek-AI

Hands-on

Yes — code/exercises

Certificate

None

• ✓Why parallel drafting loses acceptance rate as the proposal length grows in speculative decoding
• ✓How a semi-autoregressive draft restores token dependency without full autoregressive drafting cost
• ✓How confidence scheduling adapts verification length to survival probability and system throughput
• ✓How speculative-decoding gains are measured under real serving traffic instead of single-request benchmarks
• ✓How DSpark compares with Eagle3 and DFlash across reasoning, coding and chat benchmarks
• ✓How to train your own draft model with the released DeepSpec data and training pipeline
• ✓Which target models and draft sizes have released checkpoints — Qwen3 and Gemma at 4B, 8B and 14B

• •Production numbers, not lab numbers: 60-85% per-user speedup measured inside the DeepSeek-V4 serving system under real traffic
• •The full training pipeline, evaluation scripts and Hugging Face checkpoints are released — the DeepSpec repo has about 7,100 stars and 666 forks
• •Implements and evaluates three draft algorithms side by side (DSpark, DFlash, Eagle3), so it doubles as a comparison baseline
• •Benchmarked across nine suites including AIME25, LiveCodeBench and Arena-Hard-v2, covering reasoning, code and open-ended chat
• •Tackles the throughput-versus-latency trade-off explicitly, which most speculative-decoding work leaves to the serving stack

This page was written from 2 sources, 1 on domains other than arxiv.org.

1. 1.arxiv.org — 2607.05147vendor
2. 2.github.com — DeepSpec