ALoDLM

ALoDLM (Adaptively Looped Diffusion Language Models) is a family of diffusion language models from Amazon, released as ALoDLM-1.7B and ALoDLM-8B. They are initialized from the corresponding Qwen3 backbones and generate text and code by repeatedly applying shared transformer layers, keeping the latent states of unresolved tokens and refining them over recurrent passes.

Tokens that are confidently predicted are committed early and serve as context for the remaining positions. A learned halting policy decides when to stop the recurrent loop, and inference-time thresholds (a gate threshold q and an entropy threshold τ) trade off recurrent refinement against parallel token commitment. The paper reports comparisons against autoregressive and diffusion baselines on eleven reasoning, knowledge, math, science and code benchmarks, and throughput measurements for the 8B model on a single NVIDIA B200 GPU.

Weights are published on Hugging Face under CC BY-NC 4.0, and the model card states the intended use is noncommercial research on diffusion language modeling, adaptive computation, math reasoning and code generation. The card notes the benchmarks do not establish readiness for autonomous or high-consequence use. Throughput results use a separate parallel-commitment mode and do not cover 1.7B or the portable decoder.