Moonshot AI is shifting the focus of its Kimi models from coding toward biology, chemistry and pharmaceutical research. The Beijing company says gains on programming benchmarks are flattening and argues that scientific work offers harder problems where model improvements could have greater practical value.
One early integration places Kimi K3.1 Pro inside DP Technology’s BioMaster research assistant. In a reported test, the system analyzed gene data from 3,696 mouse pancreatic cells and identified 100 candidate genes in 90 minutes, compared with an estimated four hours for a person. That measures speed on one workflow, not whether the candidates lead to a successful biological discovery.
The limits remain substantial. On TCSExam, a set of 100 theoretical computer-science problems, leading models produced proofs that could be checked but did not solve any open conjecture independently. A June Citi survey of 50 US research leaders found that 72% already used AI broadly, while only 24% expected it to materially improve drug-development success rates.
The pivot therefore changes Moonshot’s target rather than demonstrating autonomous science. Models may accelerate literature work, data analysis and candidate generation, but laboratory validation and expert judgment still decide whether an output is useful. Kimi’s progress will be easier to assess through reproduced research outcomes than through another set of vendor-selected benchmark scores.