HIT and Huawei propose Dynamic-dLLM, a training-free acceleration framework with 4.48x speedup

提速4.48倍！哈工大华为新框架让扩散大模型精度无损、推理起飞

HIT Shenzhen, Huawei, and Shenzhen Hetao College proposed Dynamic-dLLM, a training-free dLLM acceleration framework that raises LLaDA-8B-Instruct throughput on GSM8k from 8.32 TPS to 37.29 TPS with almost no accuracy loss.

Why it matters: HKR-H/K/R all pass: the 4.48x speedup is clickable, and GSM8k TPS figures add concrete substance. It is inference-optimization research, not a mainstream model launch, so it fits the 78–84 band.