Moonshot AI is steering its Kimi models away from coding and toward scientific research, saying the gains in code have flattened. The Beijing lab’s international director made the case at the Web Summit in Lisbon last week, arguing that biology, chemistry and pharma demand smarter models.
Why Science, Why Now
Giacomo Ficari, Moonshot’s director of international, said science demands smarter models than law, finance or code. Coding drove much of the recent demand for advanced models, he said, but it is approaching saturation in technological progress. Science, by contrast, still has room to grow.
The shift puts Moonshot alongside American rivals turning AI toward research. OpenAI has focused on mathematics, while Anthropic’s Claude found an enzyme system in September. Moonshot is betting on biology and chemistry instead, where it says the next intelligence gains will come.
Ficari’s framing is a direct challenge to the coding-first strategy that has driven model development for two years. Benchmarks in code were the clearest way to rank models, and every lab chased them. If the scores have stopped moving, the ranking game moves somewhere the scores are harder to game.
The First Proof Point
The company is already putting the plan into practice. Beijing startup DP Technology has integrated the Kimi K3.1 Pro model into its BioMaster research assistant, which analyses gene data for biomedical work. In one test, the system analysed gene-expression data from 3,696 mouse pancreatic cells.
BioMaster flagged 100 candidate genes for further investigation in about 90 minutes, against an estimated four hours of human work. DP Technology said Kimi K3.1 Pro scored highest in its internal evaluation of models for biological research tasks, paired with its own coding tool.
Our Read
The pivot is honest about a real problem: coding benchmarks are flattening as the leaders converge, and every lab needs a new field to prove superiority. Science is the hardest field to fake, which is both the attraction and the risk. A vendor case study is a start, not proof.
There is a commercial logic too. Scientific research is where the money is patient: drug discovery budgets dwarf coding-tool subscriptions, and a model that genuinely helps in the lab commands a premium. Moonshot is trading a crowded market for a harder one with bigger prizes.
Moonshot’s case study comes with caveats. DP Technology did not name the rival model combinations it tested, and the 90-minute result is a demonstration, not a peer-reviewed finding. Pharma is cautious too: only 24 percent of research leaders in a June survey expect AI to noticeably improve drug development success.
Why It Matters
If coding is saturating, the next contest between frontier labs will be fought in science, where the work is harder to benchmark and easier to overclaim. Moonshot’s move signals that Chinese labs see the same wall as American ones. The results will take years to judge.
Background
Kimi K3.1 Pro was released in late October with about three trillion parameters and a one-million-token context window, aimed at long coding, 3D generation and complex agent work. Moonshot has set no timetable for the science push, and no new model has been announced for it.
Source: The Decoder and South China Morning Post


0 Comments