CrystalReasoner/Qwen2.5-3B-CrysReas-RL

Hugging Face
TEXT GENERATIONConcurrency Cost:1Model Size:3.1BQuant:BF16Ctx Length:32kPublished:May 15, 2026Architecture:Transformer Warm

CrystalReasoner/Qwen2.5-3B-CrysReas-RL is a 3.09 billion parameter Qwen2.5-based large language model developed by CrystalReasoner, specifically fine-tuned for generating crystal structures from natural language instructions. This model integrates supervised fine-tuning, crystallographic and physical priors through thinking traces, and reinforcement learning to enhance the validity, stability, and property conditioning of generated crystal structures. It excels at property-conditioned crystal structure generation, providing atomic coordinates and lattice parameters based on chemical formulas and material properties.

Loading preview...

CrystalReasoner/Qwen2.5-3B-CrysReas-RL Overview

CrystalReasoner/Qwen2.5-3B-CrysReas-RL is an advanced large language model built upon the Qwen2.5-3B architecture, specifically engineered for the task of generating crystal structures. Developed by CrystalReasoner, this model stands out due to its unique training methodology, which combines supervised fine-tuning (SFT) with reinforcement learning (RL) to achieve high-quality, property-conditioned crystal structure generation.

Key Capabilities

  • End-to-end Crystal Structure Generation: Generates complete crystal structures, including lattice vectors, angles, element types, and atomic coordinates, directly from natural language instructions.
  • Reasoning with Physical Priors: Incorporates crystallographic and physical priors through "thinking traces" during generation, ensuring more chemically and physically sound outputs.
  • Reinforcement Learning for Quality: Utilizes reinforcement learning with verifiable rewards to significantly improve the validity, stability, and property conditioning of the generated structures.
  • Property-Conditioned Generation: Capable of generating structures based on specified material properties, such as chemical formula and bulk modulus, as demonstrated in the example.

Good For

  • Materials Science Research: Ideal for researchers and scientists needing to generate hypothetical or novel crystal structures based on desired properties.
  • Computational Chemistry: Useful for automating the initial stages of material design and discovery by providing structured atomic configurations.
  • Integration with Material Libraries: Provides output that can be directly converted into standard formats like pymatgen.Structure for further analysis and simulation.

Finetunes
1 models