DASD-4B-Thinking: A Compact Model for Long-CoT Reasoning
DASD-4B-Thinking, developed by Alibaba-Apsara, is a 4 billion parameter dense language model engineered for long chain-of-thought (Long-CoT) reasoning in mathematics, code generation, and scientific domains. It is post-trained from Qwen3-4B-Instruct-2507 and distilled from gpt-oss-120b using a unique distribution-aligned sequence distillation pipeline.
Key Capabilities
- Specialized Long-CoT Reasoning: Excels in complex, multi-step reasoning tasks across various technical fields.
- Extreme Data Efficiency: Achieves strong performance with only 448K training samples, significantly fewer than many larger models.
- Novel Distillation Pipeline: Introduces a new paradigm of Distribution-Aligned Sequence Distillation, incorporating Temperature-scheduled Learning, Divergence-aware Sampling, and Mixed-policy Distillation.
- Open-Source Data: The training datasets, including "Superior-Reasoning-SFT-gpt-oss-120b", are open-sourced to enable reproducibility and community contributions.
- Competitive Benchmarks: Outperforms many larger open-source models in benchmarks like AIME24, AIME25, LiveCodeBench, and GPQA-D.
Good for
- Applications requiring robust mathematical and scientific reasoning.
- Code generation tasks demanding logical thought processes.
- Deploying advanced reasoning capabilities on consumer-grade hardware due to its compact size.
- Researchers interested in data-efficient distillation methods and scalable reasoning models.
Limitations
Currently, DASD-4B-Thinking operates strictly within the text space and lacks tool integration and function calling capabilities, limiting its utility in agent-based workflows. Future iterations aim to address this by integrating features like knowledge retrieval and tool invocation.