Models

Moonshot AI redirects Kimi from coding toward scientific research

Moonshot AI is shifting Kimi toward biology, chemistry and pharma. A BioMaster test analyzed 3,696 mouse cells in 90 minutes, but benchmarks and a Citi survey show limits to autonomous discovery.

Moonshot AI pivots Kimi models from coding to science as benchmark gains flatten
Moonshot AI pivots Kimi models from coding to science as benchmark gains flattenStockSnap / pixabay
By James Calloway6 min read

Why it matters

  • Kimi K3.1 Pro analyzed gene data from 3,696 mouse pancreatic cells and flagged 100 candidate genes in 90 minutes.
  • The report estimated that the same task would take a human about four hours.
  • Giacomo Ficari said: “science demands smarter AI models than law, finance, or code.”
  • On the 100-problem TCSExam, leading models produced verifiable proofs but solved no open conjectures alone.
  • A June Citi survey of 50 US research leaders found 72% using AI broadly and 24% expecting better drug-development success rates.

In a test involving Beijing startup DP Technology’s research assistant BioMaster, Kimi K3.1 Pro analyzed gene data from 3,696 mouse pancreatic cells and flagged 100 candidate genes in 90 minutes, according to The Decoder. The report estimated that the same task would take a human about four hours.

That result gives Moonshot AI a concrete workflow to test as it redirects its Kimi language models (AI systems trained to generate and work with text) from coding toward biology, chemistry and pharmaceutical research. The company cites diminishing returns in coding, but the shift also places a heavier burden on models that must reason across incomplete scientific evidence.

Why is Moonshot AI shifting away from coding?

Moonshot AI’s international director, Giacomo Ficari, made the strategic case at the Web Summit in Lisbon. He said: “science demands smarter AI models than law, finance, or code.”

Ficari’s claim frames scientific work as a more demanding test than several professional domains, including software development, finance and law. For customers, that distinction determines whether Kimi remains a coding tool or becomes a model for research workflows.

The Decoder did not identify the coding benchmark scores that prompted the pivot. It also gives no model-by-model comparison, release timetable, budget, revenue forecast or measure of customer demand for the new research focus.

Without those details, the stated slowdown remains a company assessment rather than a documented market-wide finding. Moonshot AI may see less room for easy gains in code, but the source does not show that scientific reasoning offers an easier commercial path.

A commercial pivot toward scientific buyers would require evidence different from the benchmarks used for software development. The report names only one integration partner and supplies no contract, pricing or deployment terms. The takeaway is simple: Moonshot AI has changed direction, but the source does not quantify what it gave up.

What did the BioMaster test actually show?

DP Technology, another Beijing startup, is integrating Kimi K3.1 Pro into BioMaster. The research assistant (software that supports scientific analysis) provides the vehicle for Moonshot AI’s move from coding into research workflows.

In the reported test, Kimi analyzed gene data from 3,696 mouse pancreatic cells and flagged 100 candidate genes (sequences that may influence a biological trait) in 90 minutes. The Decoder compared that performance with an estimated four hours of human work, suggesting faster triage of candidates for researchers.

The number does not establish a scientific discovery. The source does not say whether scientists experimentally validated any candidate. It also does not say whether the genes were novel or whether another model or researcher could reproduce the result. It gives no count for correct findings or missed genes, so the speed result does not reveal the quality of the biological output.

The human comparison remains loosely defined. The report does not describe the workflow, the number of people involved, the hardware, the acceptance criteria for “flagged,” or whether the four-hour estimate included interpretation and verification. Each missing condition could change the apparent advantage.

One successful screening demonstration offers evidence of utility, not independent discovery. In plain terms, Kimi may help researchers narrow a search, while the burden of proving what the candidates mean still sits with the scientists.

Can current models make independent scientific breakthroughs?

The limits extend beyond the biomedical example. On TCSExam, leading models produced verifiable proofs but could not crack any open conjectures (unresolved mathematical propositions) alone, The Decoder reported. TCSExam is a standardized evaluation (benchmark) containing 100 problems in theoretical computer science (the study of computation).

The source’s assessment was blunt: “Fully independent breakthroughs are still a long way off.” That judgment matters because Moonshot AI’s pivot rests on stronger models doing more than assist with scientific tasks.

TCSExam points to a division between formal assistance and original research. Producing a verifiable proof is not the same as selecting a previously open proposition and proving it. The result supports capability in formal reasoning, not autonomous conjecture solving.

US rivals have reached similar limits, according to The Decoder. Anthropic’s Claude found an enzyme system (a set of molecular reactions) in September, while mathematicians criticized OpenAI’s claimed mathematics and physics breakthroughs.

Those examples are not equivalent, and the source supplies little evidence with which to compare them. It gives no year for the September result, no details of the enzyme system, no named critics and no standard for deciding when a claimed solution qualifies as a breakthrough.

Verifiable proofs therefore mark progress without resolving the independence problem. The practical takeaway is that model assistance looks more achievable than autonomous scientific discovery.

Why has pharma adoption not produced equally strong expectations?

The commercial evidence is similarly mixed. In a June Citi survey of 50 US research leaders, 72 percent said they already used AI broadly, but only 24 percent expected it to noticeably improve drug-development success rates.

The gap separates adoption from expected impact. A tool can accelerate one research task without raising drug-development success rates, and the BioMaster demonstration measures workflow time rather than that outcome.

The survey shows current use without establishing a stronger forecast of improved outcomes. That distinction matters because Moonshot AI’s new focus depends on scientific value, not only broad participation in AI.

The survey cannot settle the industry’s outlook. It covers only 50 US research leaders, provides no year beyond June, and does not describe what “broadly” means or how respondents defined a noticeable improvement in success rates.

The result does not show that most research leaders oppose AI or expect no benefit. It records widespread use alongside much lower confidence in better drug-development outcomes. In plain terms, reported usage runs ahead of expected impact.

What would validate Kimi’s scientific pivot?

Moonshot AI faces two linked proof points. The first is whether Kimi can reproduce and expand the BioMaster result under controlled conditions. The second is whether stronger models can produce research findings that stand without direct human guidance.

A credible follow-up would need to preserve the 90-minute result across independent datasets, state the human baseline in reproducible terms, and track candidates through validation. It would also need to measure longer-term drug-development outcomes, not merely record adoption.

The source presents a single test, a benchmark result showing no standalone conjecture solving, and a survey that separates usage from expected impact. It does not disclose the economics of the pivot, a schedule for additional scientific releases, or plans for broader adoption.

For biology, chemistry and pharmaceutical research, the decisive issue will not be how quickly a model lists candidates. It will be whether researchers can validate those candidates, reproduce the underlying work and turn faster analysis into better drug-development results.

If those links remain unproven, the Kimi shift will be a change in product positioning rather than a demonstrated scientific lead. That is the commercial and research standard Moonshot AI still has to meet.

Original: scmp.com

Share this article:

More from James Calloway

James Calloway

Show full bio

News editor covering people, launches and day-to-day practice in artificial intelligence and machine learning at AI In Context.

722 articles

Related articles

  1. Google Publishes Co-Scientist in Nature, Opens AI Hypothesis Tool to Researchers
  2. BioBasicsBench: AI agents still get one third of routine lab analysis wrong
  3. No AI agent clears 12 percent on SemiAnalysis' new GDPval rival
  4. Kimi K3.1 Pro closes gap to Anthropic’s Claude Opus 5.5
  5. AI coding agents fully complete only 26.4 percent of real-world tasks

« Previous articleNext article »