Distribution-Aligned Sequence Distillation (DASD) is an advanced pipeline that enhances reasoning tasks like mathematical problem-solving and code generation using minimal training data. It employs techniques such as temperature-scheduled learning, divergence-aware sampling, and mixed-policy distillation. These methods allow DASD to achieve top-tier performance, even outperforming larger models like the 32B-scale versions in benchmarks such as AIME24, AIME25, LiveCodeBench, and GPQA-Diamond. Two key variants of DASD are highlighted. The DASD-4B-Thinking model is lightweight yet excels across various benchmarks, achieving notable results with multi-stage training. The DASD-30B-A3B-Thinking-Preview model uses a Mixture-of-Experts architecture, which allows for increased model capacity while maintaining efficiency through sparse expert routing. Despite being trained only on initial data, it shows impressive efficiency and quality trade-offs, particularly in AIME25 and LiveCodeBench benchmarks. The article outlines the release of several resources, including the DASD-4B-Thinking model and related datasets, which are now available on platforms like Hugging Face and ModelScope. The methodology emphasizes a systematic approach to sequence-level distillation, focusing on optimizing reasoning diversity rather than just correctness. This is achieved through a coherent pipeline that combines different mechanisms to improve both learning stability and the fidelity of reasoning transfer, making it effective within a compact model. The technical details provided include code snippets for deploying the model, indicating its accessibility for developers looking to implement these advanced reasoning capabilities.