Stable-DiffCoder is a new code diffusion large language model based on the Seed-Coder architecture. It features an innovative block diffusion continual pretraining (CPT) process that incorporates a tailored warmup and a block-wise clipped noise schedule. This model offers a systematic approach to training, aiming for stability and improved performance compared to traditional autoregressive (AR) models. The findings show that Stable-DiffCoder outperforms several existing AR and diffusion-based models, particularly in handling structured code for tasks like editing and reasoning. The model comes in two versions: Stable-DiffCoder-8B-Base, which is pretrained on model-centric code data, and Stable-DiffCoder-8B-Instruct, designed for better alignment with user instructions. Both models are available on Hugging Face for public access. The training pipeline is built to ensure a fair comparison with its AR counterparts by keeping architecture, data, and training methods consistent. This focused approach allows for a clearer understanding of how diffusion training impacts model performance, leading to actionable insights for future developments. Stable-DiffCoder has shown strong results across various coding benchmarks, including code generation and completion tasks. Through its minimal stages of CPT followed by supervised fine-tuning, it competes effectively with models around the 8 billion parameter mark and even some larger ones. For those interested in deeper insights, a Technical Report provides detailed benchmark results and analyses. The project is open for contributions and citations, encouraging researchers to explore and utilize this model further.