Aliyun's Open-Source Scientific Model LOGOS Beats Microsoft with 1/56th the Parameters
Alibaba's ATH-Token Foundry, in collaboration with the Gaoqiang Institute of Artificial Intelligence at Renmin University of China, has announced the open-source release of LOGOS. This is the first multi-domain scientific generative foundation model built on a unified scientific syntax, consistently matching or surpassing traditional domain-specific methods across six representative scientific tasks using pure sequence modeling.

What stands out is the model's exceptional parameter efficiency. With just 1 billion parameters, LOGOS-1B outperforms Microsoft's NatureLM (which has 8×7B parameters) on several key tasks.
Pioneering Unified Scientific Syntax for Heterogeneous Objects
LOGOS uses a large pre-training corpus covering seven modalities—biological macromolecules, chemical entities, and interface interactions—amounting to 44.87 billion tokens. Through a shared vocabulary, it encodes formerly heterogeneous objects like proteins and small molecules into a unified discrete token sequence.
This distinctive scientific syntax allows large models to autoregressively interpret different scientific objects within a shared generation space. It even introduces a "text description method" that lets complex spatial interaction rules be conceptualized without inputting intricate 3D coordinates, relying purely on sequence prediction.

Bridging the Gap Between Pre-training and Application
In conventional research settings, switching models between stages often requires extensive fine-tuning at deployment. LOGOS achieves strong consistency in both form and purpose, as its pre-training data sequence format matches the input and output formats of downstream tasks.
This alignment effectively removes the divide between pre-training and real-world applications, enabling direct generative capabilities without complex adaptation layers. Alibaba has now fully open-sourced the model weights, inference code, and technical report for this large model.
Related article
SAS Launches Enterprise AI Governance Tool to Manage Intelligent Entities and Mitigate Shadow AI Risks
During the recent SAS Innovate conference, analytics leader SAS unveiled a suite of governance tools designed for Agentic AI. These solutions empower organizations to create transparent, secure automated ecosystems, tackling the compliance and trust
MacWhisper 14 Update Brings Direct Transcript Editing and Fixes High CPU Usage
Jodi Bruijn, the independent developer behind MacWhisper, released version 14.0 yesterday, prioritizing enhanced transcription performance and introducing a dedicated editor view. Since its 2023 launch, MacWhisper has established itself as an AI-powe
Qwen3.5-Max Preview Tops LMArena as China's AI Models Reach Global Front Rank
On March 20, the latest rankings from the authoritative large model blind test platform LMArena revealed that Alibaba's newest flagship preview model in the Qwen series, Qwen3.5-Max-Preview, delivered exceptional performance. Scoring 1464 points, it
Related Special Topic Recommendations
Comments (0)
0/500
Alibaba's ATH-Token Foundry, in collaboration with the Gaoqiang Institute of Artificial Intelligence at Renmin University of China, has announced the open-source release of LOGOS. This is the first multi-domain scientific generative foundation model built on a unified scientific syntax, consistently matching or surpassing traditional domain-specific methods across six representative scientific tasks using pure sequence modeling.

What stands out is the model's exceptional parameter efficiency. With just 1 billion parameters, LOGOS-1B outperforms Microsoft's NatureLM (which has 8×7B parameters) on several key tasks.
Pioneering Unified Scientific Syntax for Heterogeneous Objects
LOGOS uses a large pre-training corpus covering seven modalities—biological macromolecules, chemical entities, and interface interactions—amounting to 44.87 billion tokens. Through a shared vocabulary, it encodes formerly heterogeneous objects like proteins and small molecules into a unified discrete token sequence.
This distinctive scientific syntax allows large models to autoregressively interpret different scientific objects within a shared generation space. It even introduces a "text description method" that lets complex spatial interaction rules be conceptualized without inputting intricate 3D coordinates, relying purely on sequence prediction.

Bridging the Gap Between Pre-training and Application
In conventional research settings, switching models between stages often requires extensive fine-tuning at deployment. LOGOS achieves strong consistency in both form and purpose, as its pre-training data sequence format matches the input and output formats of downstream tasks.
This alignment effectively removes the divide between pre-training and real-world applications, enabling direct generative capabilities without complex adaptation layers. Alibaba has now fully open-sourced the model weights, inference code, and technical report for this large model.
MacWhisper 14 Update Brings Direct Transcript Editing and Fixes High CPU Usage
Jodi Bruijn, the independent developer behind MacWhisper, released version 14.0 yesterday, prioritizing enhanced transcription performance and introducing a dedicated editor view. Since its 2023 launch, MacWhisper has established itself as an AI-powe
Qwen3.5-Max Preview Tops LMArena as China's AI Models Reach Global Front Rank
On March 20, the latest rankings from the authoritative large model blind test platform LMArena revealed that Alibaba's newest flagship preview model in the Qwen series, Qwen3.5-Max-Preview, delivered exceptional performance. Scoring 1464 points, it





Home






