Amazon · Looped Diffusion LLM

ALoDLM-8B

Amazon's Adaptively Looped Diffusion Language Model, in 1.7B and 8B sizes on Qwen3 backbones. Instead of giving every token the same compute, it runs shared transformer layers repeatedly: tokens it is confident about commit early, harder ones keep their latent state for more passes, and a learned halting policy decides when to stop. In the paper's 11-benchmark table the 8B averages 80.3 against Qwen3-8B's 78.5, with 87.8 on HumanEval and 49.5 on GPQA-Diamond, vendor-reported. Ships with an inference engine based on nano-vLLM. CC-BY-NC-4.0, non-commercial. Paper arXiv 2610.04198 (3 October), weights uploaded 2 October.