Amazon ALoDLM releases adaptive diffusion models

ALoDLM’s compelling idea is adaptive compute at the token level, addressing diffusion LMs’ quality gap without abandoning parallel decoding.

• –Reports 80.3 average performance across 11 benchmarks at 8B, exceeding the corresponding Qwen3 autoregressive baseline.
• –Reaches 93.25% on GSM8K at 612.4 tokens per second, roughly 2.7× faster than vLLM-served Qwen3-8B on one NVIDIA B200.
• –Includes research code and an optimized inference engine, but the speed claims depend on specialized hardware, kernels, and workload-specific settings.
• –The CC BY-NC 4.0 model license makes this valuable for experimentation while limiting straightforward commercial deployment.
• –If the approach generalizes, token-adaptive recurrence could make diffusion models more practical for latency-sensitive reasoning and coding workloads.