SDAR (Synergy of Diffusion and AutoRegression) model is a new large language model that integrates autoregressive (AR) and discrete diffusion modeling strategies. It combines the efficient training paradigm of AR models with the highly parallel inference capability of diffusion models, while delivering performance fully on par with SOTA open-source AR models. At the same time, SDAR sets a new benchmark as the most powerful diffusion language model to date. We highlight three major conclusions from our study:

[!IMPORTANT] Take-home message

Balanced Efficiency: SDAR unifies the efficient training of AR models with the parallel inference of diffusion, achieving both fast training and inference.

Fair Comparisons: In rigorously controlled experiments, SDAR achieves on-par general task performance with strong AR baselines, ensuring credibility and reproducibility.

Superior Learning Efficiency: On complex scientific reasoning tasks (e.g., GPQA, ChemBench, Physics), SDAR shows clear gains over AR models of the same scale, approaching or even exceeding leading closed-source systems.

Performance

SDAR v.s. Qwen

For SDAR models, inference hyperparameters are set to: block_length = 4, denoising_steps = 4, greedy decoding.

For Qwen3-1.7B-AR-SFT and Qwen3-30B-AR-SFT, we use greedy decoding, and the base models Qwen3-1.7B-Base and Qwen3-30B-Base are derived from the Qwen3 Technical Report.

SDAR-Sci v.s. AR Baseline

This table presents a controlled comparison between AR and SDAR under the same backbone and dataset settings. The results are averaged over 8 runs for GPQA, and over 32 runs each for AIME 2024, AIME 2025, and LiveMathBench.

SDAR-Sci v.s. Other Models

This table positions SDAR-30B-A3B-Sci(sample) against leading open-source and closed-source LLMs. Scores for external models are sourced from the InternLM/Intern-S1 repository.

JetLM/SDAR-8B-Chat-b8

SDAR

Introduction

Performance

SDAR v.s. Qwen

SDAR-Sci v.s. AR Baseline

SDAR-Sci v.s. Other Models

JetLM/
SDAR-8B-Chat-b8