Tsinghua Opensources Molecule General Model

Post LinkedIn

🔥Read original on 36氪

#chemistry-ai#biomed-model#foundation-modelbiomedgpt-molbiomedgpt-moltsinghua-airshuimu-molecular

💡Open-source chem LLM base model from Tsinghua—free tool for drug/mol AI R&D

⚡ 30-Second TL;DR

What Changed

Open-sourced by Tsinghua AIR and Shuimu Molecular

Why It Matters

Boosts accessible AI for drug discovery and materials science, enabling researchers to fine-tune for specialized chem tasks without starting from scratch.

What To Do Next

Download BioMedGPT-Mol from Hugging Face and fine-tune for custom molecule generation.

Who should care:Researchers & Academics

Key Points

•Open-sourced by Tsinghua AIR and Shuimu Molecular
•General foundation model specialized for chemical molecules
•Targets biomedicine and chemistry applications

🧠 Deep Insight

Background and context from public sources — not the original article. 6 sources cited.

🔑 Enhanced Key Takeaways

•BioMedGPT-Mol is initialized from Qwen3-8B parameters and fine-tuned for 3 epochs using a multi-task learning framework on curated public instruction datasets.[1][3]
•It achieves top performance on benchmarks like 90.4% accuracy in property prediction (surpassing Uni-Mol), 49.8% EM on chemical reactions, and 29.6% EM on description-guided generation.[1][2]
•On RetroBench, it reaches 39.1% exact match accuracy using reasoning sampling and weighted beam search, competitive with GPT-4-based methods without specialized algorithms.[2][3]

📊 Competitor Analysis▸ Show

Feature	BioMedGPT-Mol	LlaSMol	Uni-Mol (baseline)	GPT-4/Claude-3 Opus
Property Prediction Acc	90.4% (1st among LLMs) [1][2]	Lower than BioMedGPT-Mol [1]	Lower (task-specific) [1]	Outperformed [2]
Chemical Reaction EM	49.8% (best among LLMs) [1]	Lower [1]	N/A	Outperformed [2]
Description Gen EM/FTS	29.6%/77.5% [1]	EM +10.4%, FTS +15.8% lower [1]	N/A	Outperformed [2]
Multi-Prop Opt SR	95.2% [2]	Lower (GeLLM3Os lower) [2]	N/A	N/A

🛠️ Technical Deep Dive

•Initialized with pre-trained Qwen3-8B parameters and fine-tuned for 3 epochs with a learning rate (unspecified value).[1]
•Multi-task framework optimizes tasks including property prediction (e.g., ESOL/LIPO RMSE 0.945, classification 90.4%), chemical reactions (forward/retrosynthesis EM 49.8%), molecule editing (SR 74.2%), and joint property optimization (e.g., BPQ for BBBP/plogP/QED, MPQ for mutagenicity/plogP/QED, SR 95.2%).[1][2]
•Evaluated on consolidated benchmark from LlaSMol, TOMG-Bench, MuMOInstruct; uses 20-molecule beam search for optimization; 2D test-time scaling (reasoning sampling, weighted beam search) for RetroBench (39.1% EM).[1][2][3]

🔮 Future ImplicationsAI analysis grounded in cited sources

BioMedGPT-Mol enables template-free end-to-end retrosynthetic planning at SOTA levels

It achieves 39.1% EM on RetroBench comparable to GPT-4 methods without specialized search algorithms, as shown in arXiv evaluations.[2][3]

General reasoning LLMs can be efficiently adapted to molecular tasks via multi-task post-training

Fine-tuning Qwen3-8B on curated datasets yields superior performance over chemistry-specific LLMs on understanding and generation benchmarks.[1][3]

⏳ Timeline

2025-12

BioMedGPT-Mol paper submitted to arXiv (v1 on Dec 4)

2025-12

Paper updated to v2 on arXiv (Dec 11)

2026-03

Tsinghua AIR and Shuimu Molecular open-source BioMedGPT-Mol

📎 Sources (6)

Factual claims are grounded in the sources below. Forward-looking analysis is AI-generated interpretation.

🔥Read original article on 36氪

📰

Weekly AI Recap

Read this week's curated digest of top AI events →

👉Related Updates

Same topic

Explore #chemistry-ai

Same product