OpenDFM/RetroDFM-R-8B

TEXT GENERATIONPricing:Input $0.468 / Output $1.82Concurrent Unit Cost:1Model Size:8BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Jan 12, 2026License:apache-2.0Architecture:Transformer Open Weights Featherless Exclusive Cold

OpenDFM/RetroDFM-R-8B is an 8 billion parameter large language model developed by OpenDFM, specifically designed for chemical retrosynthesis prediction. It employs a reasoning-driven approach, integrating large-scale reinforcement learning with rule-based and generative reward models. This model excels at predicting reactants for target molecules, offering enhanced generalization, reliability, and interpretability compared to traditional methods. It is optimized for reconstructing multi-step chemical routes and providing human-interpretable insights for retrosynthesis planning.

Loading preview...

RetroDFM-R: Reasoning-Driven Retrosynthesis

RetroDFM-R-8B is an 8 billion parameter large language model from OpenDFM, engineered for chemical retrosynthesis prediction. Unlike conventional graph or sequence models, it leverages a unique reinforcement learning framework with both rule-based and generative reward models to achieve superior performance.

Key Capabilities & Features

  • Reasoning-Driven Prediction: Incorporates explicit reasoning processes, leading to more reliable and interpretable retrosynthesis predictions.
  • Enhanced Generalization: Demonstrates stronger generalization capabilities across diverse chemical reactions.
  • High Prediction Reliability: Outperforms existing state-of-the-art methods on standard benchmarks, confirmed by double-blind human assessments for chemical plausibility.
  • Multi-step Route Reconstruction: Successfully reconstructs complex multi-step synthesis routes for real drug molecules and materials.
  • Interpretable Insights: Provides clear, human-interpretable insights into the retrosynthesis planning process.

Training Methodology

The model undergoes a three-stage training pipeline:

  1. Continual Pretraining: Focused on retrosynthesis-specific chemical data.
  2. Supervised Fine-tuning: Utilizes distilled chain-of-thought reasoning samples.
  3. Reinforcement Learning: Further refines step-by-step reasoning and prediction quality.

Important Considerations

  • The current version may generate incorrect or misleading information; results should be verified by domain experts.
  • For SMILES notation in inputs, it is recommended to preprocess using rdkit for canonicalization.

Finetunes
1 models