AivexaNewsSearch
AI news for builders and product teamsChecked every hour

Moonshot AI pivots Kimi models from coding to science as benchmark gains flatten

Collected Nov 16, 2026

Moonshot AI is redirecting its Kimi language models away from coding and toward biology, chemistry and pharmaceutical research, citing diminishing returns on coding benchmarks. Giacomo Ficari, the Beijing startup's international director, announced the shift at the Web Summit in Lisbon, arguing that scientific work demands more capable models than law, finance or code.

The pivot already has a concrete integration. DP Technology, a Beijing startup, is putting Kimi K3.1 Pro inside BioMaster, its research assistant. In one disclosed test, the system examined gene data from 3,696 mouse pancreatic cells and flagged 100 candidate genes in 90 minutes, against an estimated four hours of human effort for the same task. That is a rough efficiency comparison rather than a validated scientific result, but it shows the intended shape of the work: high-volume candidate screening where a model can compress a researcher's first pass.

The framing matters as much as the models. Ficari's claim is that science is a harder test of intelligence than the domains vendors have competed on, which reads as a repositioning away from saturated coding leaderboards. Benchmarks tend to converge once labs optimize against them, so a pivot can signal that further gains are expensive or less legible to buyers.

Independent discovery remains distant. On TCSExam, a 100-problem theoretical computer science benchmark, leading models could produce verifiable proofs but did not solve any open conjectures on their own. US labs show comparable limits: Anthropic's Claude identified an enzyme system in September, while OpenAI's claimed math and physics breakthroughs drew criticism from mathematicians.

Demand-side skepticism is also visible. A June Citi survey of 50 US research leaders found 72 percent using AI broadly, yet only 24 percent expected it to noticeably raise drug development success rates. In practice, that gap suggests adoption is running ahead of confidence in outcomes, and pharma buyers may treat assistant-style tools as productivity aids rather than discovery engines.

Why it matters: developers building for life sciences now have another serious model family pitching at their domain, with a named integration to evaluate. Product teams should read the 90-minute screening result as workflow compression, not validation, and expect customers to ask for proof at the stage where drug programs actually succeed. A likely inference is that competition will move from benchmark scores to audited, domain-specific evaluations.

Read at The Decoder

Based on reporting from the original publisher. Visit the source for full context and later updates.

Publisher excerpt

Moonshot AI is done chasing coding benchmarks, it says. The Beijing startup is pivoting its Kimi language models toward biology, chemistry, and pharma research, arguing that science demands smarter AI than law, finance, or code. The article Moonshot AI pivots Kimi models from coding to science as benchmark gains flatten appeared first on The Decoder .