OzymandisLi/ChemCraft-7B

Related Repos: GitHub | Paper | Dataset      Authors: Peking University | IDEA Research

ChemCraft is a family of specialized open models designed for chemistry and molecular discovery. Built on reasoning-guided reinforcement learning, ChemCraft models are engineered to function as highly capable Chemical Agents. This release includes the ChemCraft-7B variant, optimized for complex chemical reasoning, molecular design, and laboratory automation tasks. With its specialized architecture, ChemCraft is well-suited for scientific research, serving as both a robust chemical assistant and a foundational backbone for downstream fine-tuning in domain-specific tasks. By integrating chemical-specific constraints into the learning loop, ChemCraft bridges the gap between general linguistic intelligence and rigorous scientific accuracy.

ChemCraft contains key capability and domain-specific advancements:

• First Agent-RL Framework for Chemistry – As the first large-scale model in the chemical domain to utilize an agent-centric reinforcement learning framework, ChemCraft aligns its reasoning processes with chemical laws. This makes it an ideal assistant for multi-step chemical tasks and a versatile base for specialized fine-tuning.

• Open-Sourced Chemical Agent Sandbox – We introduce a comprehensive, open-source sandbox for chemical agents. This environment covers a diverse range of tasks, from molecular property prediction to de novo molecule design, providing a standardized playground for evaluating and training autonomous scientific agents.

• State-of-the-Art Chemical Intelligence – ChemCraft demonstrates exceptional performance across four core chemical benchmarks. It significantly outperforms general-purpose frontier models such as Gemini and Claude, as well as existing specialized multi-agent systems, setting a new standard for AI in chemistry.

Citation for ChemCraft

If you find our work helpful, please cite it as follows:

@article{li2026agentic,
  title={Agentic reinforcement learning empowers next-generation chemical language models for molecular design and synthesis},
  author={Li, Hao and Cao, He and Peng, Shenyao and Liu, Zijing and Feng, Bin and Wang, Yu and Yan, Zhiyuan and Tian, Yonghong and Li, Yu and Yuan, Li},
  journal={arXiv preprint arXiv:2601.17687},
  year={2026}
}