Alibaba and Renmin University open-source LOGOS, a multi-domain scientific AI model.

ME News reports that on June 18 (UTC+8), according to monitoring by Beating, Alibaba’s ATH-Token Foundry, in collaboration with the Huiyan School of Artificial Intelligence at Renmin University of China, has open-sourced LOGOS, a large multimodal scientific generative model. Unlike previous approaches that train separate expert models for structure prediction and molecular generation, LOGOS is the first to achieve native unified modeling and generation of proteins, small molecules, materials, and chemical reactions within a single LLM architecture. The key breakthrough lies in encoding three-dimensional spatial interaction patterns as token sequences, enabling LOGOS to capture spatial interactions without requiring 3D coordinates—thereby eliminating bias between pretraining and downstream tasks. The pretraining corpus comprises 44.87 billion tokens, fully covering seven scientific modalities including proteins, small molecules, and chemical reactions. In six benchmark tasks, the 1B-parameter LOGOS-1B model outperformed the 56B-parameter NatureLM across multiple metrics, using only 1/56th of the parameters. In the critical AI-driven drug discovery task of pocket-ligand generation, LOGOS achieved the first-ever victory over diffusion models relying on 3D coordinates, using a purely sequence-based approach. It also attained a Top-1 accuracy of 74.8% in retrosynthesis prediction. The full model weights, inference code, and academic paper for LOGOS are now openly available. (Source: BlockBeats)