Launch of Genos, an Open-Source Genomic Foundation Model with 10 Billion Parameters

The 20th Annual Meeting of the International Conference on Genomics (ICG-20) kicked off in Hangzhou on October 23, 2025. YANG Huanming (Academician of the Chinese Academy of Sciences) and Lars Bolund (Honorary Professor at the Chinese Academy of Sciences and the Chinese Academy of Medical Sciences) delivered opening remarks on behalf of the presidium. TONG Guili (Secretary of the Leading Party Members' Group of the Department of Science and Technology of Zhejiang Province and Secretary of the Party Committee of Zhejiang Lab (ZJ Lab)) attended and addressed the meeting. Under the theme "The Future of Omics and AI", the ICG-20 brought together over 100 top-notch scientists and renowned experts from all over the world.

At the meeting, ZJ Lab and BGI-Research jointly launched Genos, the world's first open-source Human-Centric genomic foundation model with 10 billion parameters based on the 021 Science Foundation Model. The model, derived from deep optimization of human genomes, is designed to analyze ultra-long contexts of up to one million base pairs while achieving single-base resolution.

The launch of Genos marks a pivotal transition in genome research - from reading base sequences to decoding life's underlying logic. This breakthrough is poised to revolutionize clinical disease diagnosis, gene drug development, individual genomic interpretation, and life science research.

A genomic model really productive in breadth, depth and efficiency

To decipher the "book of life", we first need a complete "dictionary".

Most existing AI models for genomics are trained on just a few reference genomes, failing to capture the vast diversity of human genetics. Genos breaks this fundamental limitation, being trained on a comprehensive set of 636 "telomere-to-telomere" (T2T) human genomes from diverse global populations that systematically incorporate high-quality data from the Human Pangenome Reference Consortium (HPRC) and the Human Genome Structural Variation Consortium (HGSVC) – aiming to minimize data bias at its source and offer a more holistic representation of human genetic diversity.

Genomic language is extremely complicated, since a tiny single-base mutation's impact may be governed by "distant" regulatory elements located several megabases away. This demands a model that combines both "microscope-level" single-base precision and "wide-angle-lens-level" comprehension of megabase-scale ultra-long contexts.

Much like reading an epic literary work – where one can not only remember every precise detail but also grasp the overarching narrative - Genos solves this challenge through a Mixture-of-Experts (MoE) architecture. This innovative design functions as an elite team of specialists, with its on-demand activation mechanism allowing Genos to maintain a 10-billion-parameter knowledge base for a given task while significantly reducing computational costs and resource consumption compared to models of equivalent scale, thus making it truly "powerful and user-friendly".

In order to fully verify model performance, the research team subjected Genos to a series of tests. In classical tests for genomic element identification, long-range regulatory element prediction, and mutation pathogenicity prediction, Genos outperformed similar models. Particularly in long-sequence tests, it demonstrated exceptional contextual analysis by effectively deciphering genomic "dark matter". Across all core evaluation metrics, Genos achieved superior performance compared to current global models, proving its unparalleled comprehensive capabilities in this field.Since joint launch, Genos has been open-sourced, available in two versions - 1.2 billion (1.2B) and 10 billion (10B) parameters that cater to diverse application scenarios. Both versions' weights, architecture details and complete training pipelines have been posted on global open-source platforms like GitHub, Hugging Face and ModelScope under MIT Open Source License, enabling open access for researchers worldwide. Concurrently, Genos has been deployed on ZJ Lab's zero2x (a digital infrastructure for open science) and BGI's DCS Cloud.

Fostering interdisciplinary integration between liberal arts and sciences to extend human creativity with "AI+"

Genos was born from a bold cross-disciplinary integration and talent cultivation initiative. The core research team behind Genos emerged from the Foundation Model Training Program jointly launched by ZJ Lab and BGI-Research. This program engaged 50 seed fellows in AI-powered genetic model training, achieving interdisciplinary collaboration. Through six months of intensive face-to-face research, the team completed the first stage of Genos development and training, accelerating innovations through exceptional talent density and research intensity.

Unlike human language, genomic language—base sequence—represents an objective reality that has evolved over tens of millions of years without any human intervention or priori knowledge. In this sense, Genos constitutes an AI model purely trained through unsupervised learning on genomic data—a process that reconstructs the mysteries of life.

The collaborative team's research demonstrated that genome is the key to Genos training. Just as WANG Jian, Academician of the Chinese Academy of Engineering and Director of ZJ Lab, pointed out, "Tokenization of all scientific data represents the pivotal approach for humans to transcend their own limitations and make sense of the world through AI technologies, thereby unlocking new scientific laws."

Beyond textual data, scientists require diverse inputs including gene sequences, protein structures, chemical molecules, and spectra. These fundamentally differ from human language, presenting significant challenges for integration into foundation model frameworks.

ZJ Lab is tackling this challenge head-on by pioneering 021 Science Foundation Model from scratch - designed to enable deep scientific reasoning and autonomous knowledge discovery, thereby advancing "AI + Science & Technology".

As a domain-specific model built upon 021 Science Foundation Model, Genos extends inputs from linguistic space to scientific modal space. This expansion, combined with efficient encoding of human genomes, endows Genos with distinctive capabilities. In the case of prediction of diseases caused by genetic mutations, test results show that Genos not only processes raw DNA sequences but also harnesses 021 Science Foundation Model's reasoning capacity to achieve state-of-the-art performance in generating biologically coherent explanations and predictions.

"AI is not simply a revolution in technological tools; it is a tool to revolutionize scientific fields." WANG Jian, Academician of the Chinese Academy of Engineering and Director of ZJ Lab, highlighted in his keynote speech: "We have entered the era of the Third Paradigm that is compute-intensive, data-driven and model-based, also called computing-driven scientific revolution. The integration of AI with science and technology can significantly expand human creativity."