Moonshot AI redirects Kimi models from coding to science research

Moonshot AI, the Beijing startup behind the Kimi language models, is redirecting those models toward biology, chemistry and pharma research. The stated reason is that the company sees diminishing returns in coding. International director Giacomo Ficari said at the Web Summit in Lisbon that science demands smarter AI models than law, finance or code. The article reports this as his statement, not as a direct quotation.
A first application is already under way. Another Beijing startup, DP Technology, is integrating Kimi K3.1 Pro into its research assistant BioMaster. In one test, the AI analyzed gene data from 3,696 mouse pancreatic cells and flagged 100 candidate genes in 90 minutes. The article compares that with an estimated four hours of human work for the same job.
The article stresses that fully independent AI breakthroughs are still a long way off. On TCSExam, a 100-problem theoretical computer science benchmark, leading models produced verifiable proofs but could not crack any open conjectures alone. The article says US rivals hit similar limits: Anthropic's Claude found an enzyme system in September, but mathematicians criticized OpenAI's claimed math and physics breakthroughs.
The pharma industry is not sold yet either. A June Citi survey of 50 US research leaders found that 72 percent were already using AI broadly, but only 24 percent expected it to noticeably improve drug development success rates.
Key facts
- Moonshot AI, a Beijing startup, is redirecting its Kimi language models toward biology, chemistry and pharma research because it sees diminishing returns in coding.
- International director Giacomo Ficari said at the Web Summit in Lisbon that science demands smarter AI models than law, finance or code.
- DP Technology is integrating Kimi K3.1 Pro into its research assistant BioMaster; in one test the AI analyzed gene data from 3,696 mouse pancreatic cells and flagged 100 candidate genes in 90 minutes, against an estimated four hours of human work.
- On TCSExam, a 100-problem theoretical CS benchmark, leading models produced verifiable proofs but could not crack any open conjectures alone.
- A June Citi survey of 50 US research leaders found 72 percent already using AI broadly, but only 24 percent expecting a noticeable improvement in drug development success rates.
Why it matters
A Chinese model developer is publicly saying that coding, a major battleground for AI labs, is yielding diminishing returns for it, and that it wants to aim its Kimi models at scientific research instead. Ficari's argument is that science needs smarter models than law, finance or code. The article treats this as a sign that US rivals face similar limits, pointing to Claude's enzyme system find in September and the criticism of OpenAI's claimed math and physics breakthroughs.
Who it affects
Researchers in biology, chemistry and pharma are the intended audience, along with teams such as DP Technology that build research tools on top of Kimi models. It also touches the AI labs competing on coding, since Moonshot AI is signalling that it sees less to gain there. Pharma decision-makers are a separate group: the Citi survey shows most research leaders use AI, but few expect it to lift drug development success rates.
How to use it
The one concrete route in the article is DP Technology's BioMaster research assistant, which is integrating Kimi K3.1 Pro. The test described involved analyzing gene data from mouse pancreatic cells and flagging candidate genes. No financial terms, funding or partnership details between Moonshot AI and DP Technology are given.
How solid is it
The strategy claim rests on Ficari's remarks at the Web Summit in Lisbon, relayed as reported speech. The only hard result is a single DP Technology test: 100 candidate genes flagged in 90 minutes from 3,696 mouse pancreatic cells. The four-hour human comparison is an estimate, and no method, error rate or validation of the 100 candidate genes is given. No benchmark scores for Kimi K3.1 Pro are given, and no figures show the flattening of coding benchmark gains. The TCSExam result describes leading models in general, not Kimi specifically.
Risks and caveats
The article itself says fully independent breakthroughs are still a long way off: on TCSExam, leading models could not crack any open conjectures alone. It does not say Moonshot AI is abandoning coding entirely; it speaks of redirecting and diminishing returns. Skepticism also shows up in the market: only 24 percent of the 50 US research leaders in Citi's June survey expect AI to noticeably improve drug development success rates, even though 72 percent already use AI broadly. That is a small survey sample.