Alibaba open-sources LOGOS, a 1B-param science model that beats Microsoft NatureLM on multiple tasks
阿里开源首个统一科学大模型 LOGOS,仅用 1/56 参数超越微软 NatureLM
Alibaba's ATH-Token Foundry and Renmin University's Gaoling School of AI open-sourced LOGOS, a generative model that uses a unified 'science grammar' to handle seven modalities including proteins, small molecules, and materials. It encodes 3D pocket-ligand contacts as discrete tokens, predicting spatial interactions without explicit 3D coordinates. LOGOS-1B uses only 1/56 the parameters of Microsoft NatureLM (8×7B) and matches or beats domain-specific methods across six science tasks. Pretrained on 44.87B tokens, it shares the same sequence format and next-token prediction objective for both pretraining and downstream tasks, eliminating heavy adaptation. Weights, inference code, and the tech report are fully open on HuggingFace and GitHub.
Why it matters: Alibaba and Renmin University's Gaoling open-sourced LOGOS, which turns seven kinds of scientific data into token sequences and beats Microsoft's NatureLM with 1/56 the parameters.