NederlandsChinese company releases free AI model for chemistry, trained on its own lab work
The Chinese company Deep Principle has released Hydrogen-1.0, an AI model for chemistry and materials research that anyone can download. It also learned from failed experiments in the company's own lab. The results have so far been measured by the company itself.
The Chinese AI company Deep Principle, based in Hangzhou, released Hydrogen-1.0 on 2 November. It is an AI model that helps researchers with chemistry, materials science and biological research. Co-founder and chief technology officer Chenru Duan announced it on X.
What makes it special is that anyone may download and use the model, including for commercial purposes. The so-called weights, the numbers that hold everything the model has learned, are available under a free licence (MIT) on Hugging Face and on the Chinese platform ModelScope.
Learning from real experiments, including failed ones
Deep Principle built Hydrogen-1.0 on GLM-5.3, an existing open language model from the Chinese company Z.ai. The company trained that model further with expert knowledge from chemistry, materials science and life sciences. It also used the digital lab notebooks of its own laboratory: reports of real experiments, including the attempts that failed. This way the model learns not only what works, but also what goes wrong along the way.
The model is large: 753 billion parameters, the adjustable numbers of an AI model. It can read a text of about one million tokens (word parts) at once, enough for a stack of scientific papers. According to the makers, it runs on a server with eight powerful graphics cards. There are separate, lighter versions for the newest chips from Nvidia and AMD.
What can it do?
On FrontierScience-Research, a test with open research questions, Hydrogen-1.0 scores 38.3 per cent, according to the company. OpenAI's GPT-6-Astra scores 33.3 per cent. On a well-known test of difficult multiple-choice questions about science, it performs about as well as the Chinese model Kimi-K3.
In Deep Principle's own tests, the model does better than GPT-6-Astra at scientific calculations and at predicting what is formed in a chemical reaction. At working out the structure of a molecule from a measurement (NMR), GPT-6-Astra is better.
The company also had the model come up with new metal compounds with a desired electrical property, in five rounds. According to the company, Hydrogen-1.0 did better than GPT-6-Astra and Kimi-K3 here.
In examples, Deep Principle shows how the model writes a computer simulation of molten plastic, finds a mistake in it and fixes it. It also wrote a work plan for a laboratory robot. Before such a plan is carried out for real, the model first plays it through virtually. In doing so, it found that the robot wanted to draw up liquid without picking up a clean pipette tip, and it improved the plan. This way, mistakes are found before expensive substances are lost.
What still has to be shown
All the figures come from Deep Principle itself, and the company chose the tests itself. It does not share the training data or the exact training method. The next step is for researchers outside the company to test the model themselves. That is really possible, because the weights are freely available.
What does this mean for the road to the singularity?
AI is increasingly helping with scientific discoveries. Deep Principle itself mentions the mathematical proof about flow equations that OpenAI reported, and the discovery of a new enzyme system by Anthropic. Those results came from closed models, which outsiders cannot run or examine themselves.
With Hydrogen-1.0, universities and smaller labs can use a scientific AI model on their own computers, check it and improve it further. If learning from real lab experiments turns out to work well, the work between idea, experiment and improvement can go faster. That is exactly the kind of speed-up with which AI can raise the pace of research into new materials and medicines.
What is the technological singularity? Read the explanation →
