Insilico Medicine, a U.S.-based, Hong Kong-listed artificial intelligence drug discovery company, is launching a new service that will train general-purpose large-scale language models (such as OpenAI’s GPT or Alibaba’s Qwen) to handle biology and chemistry tasks.
Insilico founder and CEO Alex Zhavoronkov says generalist models are “failing miserably” on benchmarks used to measure how well AI performs scientific tasks wealth. “You test it five times on the same task and you’ll see that it’s nowhere near state-of-the-art…it’s basically worse than random. This is complete rubbish. “
Even better are specialized AI models trained directly on chemical or biological data. But these models often do not allow users to prompt in simple language (as with general-purpose models), and they also lack the ability to complete tasks beyond specialized scientific functions.
Insilico’s new “Scientific MMAI Gym” is designed to train general-purpose large-scale language models into ones that can perform as well as specialized models.
The gym is a hub for Insilico, part of what the company calls its “long-term roadmap toward pharmaceutical superintelligence.” The startup is part of a group of biotech companies trying to use machine learning and artificial intelligence to research and design new drugs. But with Gym, Insilico is now targeting other biotech and pharmaceutical companies, offering them the ability to train new AI models.
Insilico will “train” the model using a combination of domain-specific datasets, reward models, and reinforcement learning, and claims this process can improve model performance by up to 10x relative to key benchmarks in chemistry and biology, even approaching the performance of models specifically designed for these scientific tasks.
But why would a company decide to train a general model instead of using a specialized model? The reason is flexibility: a specialized model is very good at doing one thing, like drug discovery, but not anything else; in contrast, a trained generalist model, even if it doesn’t exactly match the performance of the expert model, can maintain its ability to perform many other tasks. This means startups can only rely on one large model, rather than a series of specialized models.
“If the model is small, it starts to forget some of the more primitive tasks for which it was designed,” Zavoronkov said. “If the model was large, you wouldn’t have this problem.”
Zhavoronkov admitted that even the universal model through the Insilico “gym” still does not perform as well as the most advanced professional models. “In order for them to reason from molecular simulations, they need to understand and observe the physics. The language isn’t really designed for that, so they’re going to be a little bit worse compared to cutting-edge physics-based models,” he explains, although he expects this to improve in the coming years.
However, as LL.M.s become more common and as more startups adopt them, Zhavoronkov said he hopes Insilico becomes “the number one trainer for these models.” He said Insilico has had conversations with potential clients about training programs; while he did not name specifics, he said he had contacted “top cutting-edge players in the United States.”
Insilico, Hong Kong and biotech
Founded in 2014, Insilico is working to become one of the first startups to bring a drug designed entirely by artificial intelligence to market through clinical trials. One of the startup’s main efforts is to develop a drug to treat idiopathic pulmonary fibrosis, a disease that causes scar tissue to form in the lungs, making it difficult to breathe. The startup says it has successfully moved its drug into clinical trials only 18 monthsfar shorter than the average four years for more traditional biotech companies. Last year, the drug Completed Phase II clinical trialthe researchers concluded that the results warranted “further study in larger, longer clinical trials.”
Insilico also targets other diseases, such as inflammatory bowel disease, as well as researching new cancer and GLP-1 drugs.
In December, Insilico raised HK$2.3 billion ($295 million) in an IPO, the largest biotech initial public offering in a Chinese city in 2025. The IPO attracted the following companies Eli Lilly and CompanyTencent and Oaktree Capital are cornerstone investors.
The startup’s share price has soared since its first trade on the Hong Kong Stock Exchange on December 30. As of January 16, Insilico’s shares were trading at HK$54.75 (US$7.02), currently worth more than double the IPO price of HK$24.05 (US$3.08).
The Hang Seng Biotech Index, which tracks the 30 largest biotech companies listed in Hong Kong, is up 100% in the past 12 months, well ahead of the benchmark Hang Seng Index’s 37% gain.
Insilico isn’t the only Hong Kong-listed artificial intelligence startup whose shares have soared in recent weeks. Shares of Minimax, a Chinese consumer artificial intelligence startup, have risen 160% since its listing on January 9. The share price of chip design company Biren has also risen by more than 90% from the IPO price.
Still, both U.S. and Chinese investors are wondering whether the AI boom can last. While Zavoronkov is closely watching the possibility of an AI bubble forming in the stock market, he is optimistic that AI drug discovery will be safer than other industries from bursting. “People can live without conversational assistants or AI-generated movies. But they can’t live without drugs.”