Moonshot AI is redirecting its Kimi language models toward scientific research, including biology, chemistry, and pharmaceutical work. The move follows diminishing returns in coding, where further benchmark gains are becoming harder to achieve, according to the company’s international director, Giacomo Ficari.
Scientific work presents a harder test
Speaking at the Web Summit in Lisbon, Ficari said science demands smarter AI models than law, finance, or code. That distinction matters: helping with a defined task, such as analyzing data, is different from producing a result that advances scientific understanding on its own.
The shift puts Kimi in areas where AI could assist researchers with complex material and time-consuming analysis. The source describes a direction for the models, rather than evidence that they can independently conduct research from start to finish.
A research assistant puts Kimi to work
Beijing startup DP Technology is integrating Kimi K3.1 Pro into its research assistant BioMaster. One reported test involved gene data from 3,696 mouse pancreatic cells. The AI flagged 100 candidate genes in 90 minutes, compared with an estimated four hours of human work.
The result illustrates a practical role for AI in research: helping process data and identify candidates for further attention. The reported comparison concerns the time taken for that analysis. It does not establish that the flagged genes are confirmed discoveries or that the model can replace researchers who assess and validate results.
For scientific teams, that distinction shapes how such tools may be useful. A faster first pass could help researchers focus their effort, while the findings still require human scrutiny. The example shows a task the system performed; it does not demonstrate a general ability to solve scientific problems independently.
Independent breakthroughs remain out of reach
Evidence in theoretical computer science points to a limit. On TCSExam, a 100-problem theoretical CS benchmark, leading models produced verifiable proofs but could not crack any open conjectures alone.
The source says US rivals have hit similar limits. It also describes contrasting reactions to claims of scientific progress: Anthropic's Claude found an enzyme system in September, while mathematicians criticized OpenAI's claimed math and physics breakthroughs. These examples underline why claims about AI discoveries can be difficult to assess and why proof of a result is not the same as independent problem-solving.
Drug development promises are still uncertain
Pharmaceutical research is another target for AI, but industry expectations remain cautious. A June Citi survey of 50 US research leaders found 72 percent already using AI broadly. Only 24 percent expected it to noticeably improve drug development success rates.
That gap separates current adoption from confidence in a major outcome. Broad use suggests AI has a place in research workflows, but the survey figures do not show that AI has already improved the chances of developing successful drugs.
Moonshot AI’s pivot reflects both opportunity and constraint. Kimi may help researchers handle specific analytical work, as the BioMaster test suggests, while the benchmark and survey findings leave larger questions open. The challenge is to show that assistance with individual tasks can translate into reliable scientific progress.