← All models

DASD-4B-Thinking

Alibaba · 2026-01-14 · 4B parameters

DASD-4B-Thinking is Alibaba Cloud's open-weight 4B dense long chain-of-thought reasoning model, released on January 14, 2026 and post-trained from Qwen3-4B-Instruct-2507 by distillation from gpt-oss-120b through a distribution-aligned sequence distillation pipeline that uses 448K training samples. The 4,022,468,096-parameter checkpoint keeps Qwen3's 36-layer GQA architecture with tied embeddings and is served with a 262,144-token context. It targets mathematics, code generation, and scientific reasoning, lacks tool-calling support, and is released under Apache 2.0.

Benchmark scores