Valued at more than a billion dollars, the start-up develops AI models capable of reasoning in biology.
AI will profoundly revolutionize the healthcare sector. This is one of the most repeated promises in Silicon Valley. Owkin has been betting on it for ten years. Launched in 2016 in Paris, the Franco-American unicorn develops AI models specialized in the discovery of new molecules and diagnostic assistance. Its long-term ambition is to build a general AI expert in biology, capable of formulating scientific hypotheses and validating them independently. And the timing seems to prove him right. At the start of 2026, AI in health is accelerating like never before: health is the leading sector of application of AI in France according to France Digitale, and the first specialized AI agents are arriving on the market. Owkin, with more than 300 million dollars raised and a valuation of one billion, is the French standard-bearer of this wave.
Owkin 0: an LLM trained to reason in biology
The start-up began by developing a language model trained to reason in biology. Called Owkin 0 and unveiled last August, the model was designed to allow researchers to identify new therapeutic targets (the proteins or genes on which a future drug could act). The model was trained, using reinforcement learning (RL), using patient data from Mosaïc, a consortium founded by Owkin which brings together ten hospitals spread between France, Switzerland, Germany and the United States. “It’s a unique dataset in the world, in which we characterized patients in a very, very deep way, in which we generated a lot of genomic, transcriptomic, medical imaging and clinical data,” explains Eric Durand, VP data sciences at Owkin.
As a result, on therapeutic target discovery benchmarks, Owkin 0 outperforms the frontier models of Google (Gemini) or OpenAI (GPT). However, the model comes up against catastrophic forgetting, a phenomenon whereby a model specialized in one task sees its performance drop on others. “When you specialize a model in one task, it tends to become less good in others,” summarizes Eric Durand. Thus, since the publication of Owkin 0, the company has internally trained a family of derived models, each specialized in a specific biological task (analysis of therapeutic targets, characterization of genes and proteins, etc.).
Owkin K: an AI agent to orchestrate research
Specialized models that work individually but show their full potential when coordinated. This is where Owkin K comes in, an AI agent capable of operating on all the models developed by Owkin. The idea is to offer an interface to researchers and pharmaceutical laboratories behind which the agent will mobilize the right specialized model, the right databases and the right analysis tools according to the query asked. “This allows non-technical personnel to carry out hyper-in-depth analyzes on very rich data,” says Eric Durand. Another advantage: interoperability. “A workflow developed by PhDs in a pharma can, thanks to a protocol like MCP, connect to a workflow developed by PhDs in Owkin, and the two work together autonomously. It’s a layer of intelligence that didn’t exist, that just wasn’t possible even two years ago,” analyzes the VP data sciences.
On the customer side, the platform targets two populations. First the academic world, Owkin has notably established a partnership with Paris-Saclay, where the tool is provided free of charge to researchers. “What Owkin gains by doing this is to learn by working with the best researchers in the world, to understand the use cases in which the tool works well, in which it is still limited,” explains Eric Durand. Then, and this is the heart of the economic model, the large pharmaceutical laboratories, which access Owkin K under an annual license as a co-pilot for the discovery of new drugs.
To accelerate the deployment of its technology stack, Owkin recently signed two strategic partnerships. First with Anthropic: via an MCP server, Owkin made a functionality called Pathology Explorer accessible directly in Claude, which allows tumor images to be analyzed by simple conversation. “Instead of needing a pathologist looking through a microscope, you are on Claude and you ask questions like: what is the grade of this tumor?”, illustrates Eric Durand. Also, with Nvidia: the partnership concerns the training of Owkin 0. The GPU manufacturer provides not computing power but its technical expertise to optimize the training of models.
The ultimate goal
Ultimately, Owkin K aims to evolve well beyond the simple co-pilot. Today, a researcher must ask iterative questions to the agent to gradually refine their analysis. Tomorrow, it will be enough to give it an overall objective and the agent will produce its own research independently. “We see that these agents are capable of being more and more autonomous, of taking on increasingly complex tasks and of working alone for longer and longer periods of time,” observes Eric Durand. It is this trajectory that leads to what Owkin calls “biological artificial super intelligence”, a system capable not only of formulating new scientific hypotheses, but also of validating them in complete autonomy.
However, a physical obstacle remains: if the AI generates a hypothesis in a few seconds, validating it in the laboratory takes much longer. “Cancer cells take, for example, 48 hours to replicate,” recalls Eric Durand. To get around this constraint, Owkin also works, in parallel, on biological simulation.
On the calendar, the R&D director refrains from any predictions. “Even the founders of DeepMind, Anthropic and OpenAI would not venture into predictions,” he insists. One thing, however, seems certain: “We are going to have this year, in 2026, important discoveries for biology made by artificial intelligence.” The race has only just begun.
