ON 9 July 2026, researchers from US universities published a paper in the top-rated journal  Science describing a powerful new AI tool for biomedical research. A year earlier, they had made the initial manuscript available as open access to all by publishing it on bioRxiv, an online repository, and making the software available on GitHub, an online platform for researchers to share code.

The tool is called Biomni, and it is the first general-purpose biomedical AI agent. It does not just answer questions. Give it a plain-language request, and Biomni searches databases, writes and runs analysis code, identifies disease-causing genes, and even generates step-by-step lab instructions. In one test, it completed a complex data analysis in 35 minutes that took a human expert three weeks. In another, researchers followed Biomni's cloning protocol exactly as written, and the experiment worked on the first attempt. Since its release in 2025, over 10,000 scientists worldwide have already adopted it. But the most important thing about Biomni is not its capability; it is that anyone can use it. The code is public. The underlying AI model is freely available. There is a simple web interface so biologists can use it without writing code. There are no licensing fees, no corporate gatekeepers, but only the shared tools without walls.

OPEN SCIENCE 

This openness was not born out of nothing. It is based on a long tradition of collaboration that has made parts of biology into a cooperative science on the planet. The Protein Data Bank started in 1971. It was the first biological database freely available to researchers anywhere. It held the 3-D structures of proteins, the molecular machines that do most of the work in our cells. The same principle applies to DNA sequence databases, gene expression data and disease-related information today. All are accessible to anyone with an internet connection, free of charge. The same spirit applies to physical materials. Scientists worldwide share plasmids, cell lines and other lab materials through non-profit organisations. Preprint servers such as bioRxiv enable scientists to share their results freely before journal peer-review, making science faster and more transparent. The open sharing of research, for example, was crucial to developing vaccines at unprecedented speed during the COVID-19 pandemic. Open biology is not an obscure ideal. It is the infrastructure that makes modern medicine happen.

AI is changing how biology is done today. It enables investigators to search millions of papers, analyse large datasets, predict protein structures and simulate how drugs work in the body. These capabilities are driving discovery at an accelerated pace. But there is a problem: Much of this AI is behind closed doors. Commercial AI tools don't reveal their inner workings. If you cannot see the code, you cannot check the results. If you cannot change the tool, you cannot adapt it to your needs. If you cannot run it on your own computer, you are at the mercy of whoever is running the servers. This is why open source AI tools are important. "Open source" means the code for the tool is available for anyone to see, use and improve. As mentioned before, one such place where these tools are available is GitHub, a repository for free access to the source code of the deposited software. Hugging Face is a similar platform for sharing AI models. By sharing tools and models like these, other scientists can check the work, spot errors, and use each other's discoveries as a springboard.

BIOMNI AND OPENNESS

Biomni has been built with the philosophy of openness. The researchers used an AI agent that read about 2,500 recent research papers from bioRxiv in 25 different areas of biology. From publicly available resources, they assembled a unified system that integrates 150 specialised tools and software packages, and 59 databases that use them. The agent itself is an AI that can reason, write and execute code on a computer. It does not rely on pre-written templates. It makes new workflows on the fly for every task it gets. Importantly, everything is kept in the open domain. The full code is on GitHub, and the AI model is on Hugging Face. Data and materials are all publicly available. Researchers can download everything, add their own improvements and do analyses on their own computers.

The results are remarkable. For rare disease diagnosis, Biomni matched expert doctors in accuracy while cutting analysis time from over 110 minutes to about 3 minutes. For identifying genes linked to diseases, it reached 80% accuracy, matching what specialists found. But it took only 4 minutes per analysis and not 90 minutes! When Biomni was given raw data from fitness trackers of over 1,000 people, which had more than 1.4 billion heart rate measurements, it autonomously identified six key health markers linked to COVID-19 and generated publication-quality charts. It analysed over 300,000 individual cells from human embryos, identifying both known and previously unrecognised genetic regulators of bone development. Most strikingly, Biomni designed a complete lab protocol for a gene-editing experiment. This included specification of all the DNA pieces needed and the instructions for assembling them. Researchers followed the protocol exactly, and the experiment worked perfectly on the first attempt.

The Biomni team does not hide the limitations. The agent still struggles with tasks that require deep scientific judgment or creative thinking. It can give wrong answers that seem reasonable, so a human scientist must always verify its work. But these limitations are precisely the reason why openness is so important. Having the code open allows the global scientific community to find errors, recommend improvements and build on each other's work. When the AI model is public, researchers can run validation studies themselves. When the data is shared, results are replicable. As one of the researchers put it: "Biomni is a powerful tool, not a decision maker." The human still asks the questions, judges the results, and decides which directions to pursue. The team sees several directions for improvement. AI models that learn from experience could get better over time. The integration of different types of data, e.g., text, images and structured data, could lead to more sophisticated reasoning. Allowing Biomni to discover and incorporate new tools automatically would keep it current. But the most important development is already happening: the global community is building on this work. Over 10,000 scientists are already using Biomni for their everyday research. Because it is open, research groups everywhere can adopt it immediately without licensing barriers. And because it is open, anyone can contribute improvements.

NEED FOR THE GLOBAL SOUTH

Biomedical research is becoming ever more inaccessible to researchers, especially those from the Global South. The tools are costly. Expertise is concentrated in wealthier institutions. The benefits go to those who can afford the access. Biomni and other open-source AI agents offer a different future. In low-income countries, laboratories are already being equipped with AI tools for research on infectious diseases, with a focus on neglected diseases affecting the Global South. If the means of discovery are public, then a researcher in a small lab with limited funding can do what used to take a team of specialists. Sharing knowledge means the global community can build on each other's work rather than duplicating it. Science benefits everyone, not just the rich, when the benefits are shared.

As Biomni's lead developer put it: "This is not about machines taking over science, but about machines becoming powerful partners to augment human researchers." That partnership, built on open principles, could be one of the great democratic projects of our time: if the tools are kept open, the code remains free, and the benefits are shared. The question is whether a profit-driven economic system will keep it that way, or whether public intervention and demand will be needed to ensure this endeavour remains part of the public commons.