Diffusion language models promise bidirectional context and infilling capabilities that autoregressive coders lack, yet practical systems remain heavyweight. We introduce CoDA, a 1.7B-parameter diffusion coder trained on TPU with a fully open-source training pipeline. CoDA pairs large-scale diffusion pre-training with code-centric mid-training and instruction tuning, enabling confidence-guided sampling that keeps inference latency competitive. On Humaneval, MBPP, and EvalPlus, CoDA-1.7B-Instruct matches or surpasses diffusion models up to 7B parameters. Our release includes model checkpoints, evaluation harnesses, and TPU training pipelines to accelerate research on lightweight diffusion-based coding assistants.
@article{arxiv.2510.03270,
title = {CoDA: Coding LM via Diffusion Adaptation},
author = {Haolin Chen and Shiyu Wang and Can Qin and Bo Pang and Zuxin Liu and Jielin Qiu and Jianguo Zhang and Yingbo Zhou and Zeyuan Chen and Ran Xu and Shelby Heinecke and Silvio Savarese and Caiming Xiong and Huan Wang and Weiran Yao},
journal= {arXiv preprint arXiv:2510.03270},
year = {2025}
}