Starts Nov 23Live Cohort 3 of Vibe Coding & AI Agents: 3 weeks, 9 live hands-on sessionsCohort 3 (Nov 23) →
ReleaseResearch Paper · For R&DThe Decoder→ original

Moonshot AI Shifts Kimi From Code to Science: Gene Analysis in 90 Minutes Instead of 4 Hours

Moonshot AI sees that models have nearly stopped improving quickly at programming, and is pivoting Kimi to biology and pharma. The BioMaster assistant, built on Kimi K3.1 Pro, analyzed data from 3,696 mouse cells in 90 minutes and flagged 100 candidate genes. A human needs about four hours for the same job. Think of it as handing the first pass through the logs to an AI while the engineer keeps the conclusions. AI is not yet making discoveries without a human.

AI-processed from The Decoder; edited by Hamidun News
Moonshot AI Shifts Kimi From Code to Science: Gene Analysis in 90 Minutes Instead of 4 Hours
Source: The Decoder. Collage: Hamidun News.
Listen to article
Article Fact Sheet & Key Figures
ParameterDetails
Primary sourceThe Decoder ↗
Editor-in-Chief
Citi survey (50 US executives)72% use AI, 24% expect it to improve success rates
TCSExam, 100 problemsmodels produced verifiable proofs, but not a single hypothesis
Moonshot director Ficariscience demands smarter models than law, finance and code

Moonshot AI, the Chinese startup behind the Kimi family of language models, has decided to step back from the race for top programming results. There, every new improvement costs more, while the gains on tests are barely noticeable. Instead, the company is turning Kimi toward biology, chemistry and pharmaceutical research.

There is already a concrete example: BioMaster, a research assistant in which Beijing-based DP Technology embedded the Kimi K3.1 Pro model, analyzed gene data from 3,696 mouse pancreas cells in 90 minutes and flagged 100 candidate genes. By estimate, a human would need about four hours for the same work.

Picture a developer whose AI sorted through a huge pile of logs in an hour and a half instead of four and picked out a hundred suspicious spots: a human still has to check them all, but the starting point arrives much faster. For those building products on LLMs, this is a signal of where the next wave of demand may shift.

Why Moonshot Is Moving Away From Code

Moonshot AI's international director, Giacomo Ficari, said at Web Summit in Lisbon that science requires smarter models than law, finance or programming. The logic: if an assistant for code or legal documents is already good enough for most tasks, further improvements deliver less and less noticeable benefit. Research work sets a higher bar, which leaves room to pull ahead of competitors.

This is not an abandonment of code as such, but a change of priority. The startup is betting that hard scientific problems will become a new proving ground and a new market for Kimi.

What BioMaster Actually Showed

It is important to understand what exactly got faster. The model did not make a discovery: it sifted a gene dataset and proposed a list of 100 candidates, which a biologist then verifies personally. That is the familiar role of an assistant: take over routine first-pass processing and shorten the path to a meaningful hypothesis. The gain, from four hours to 90 minutes, more than a twofold speedup, is nice, but it is a single test, not proof that a whole lab would work this way. The human time here is an estimate, not a measurement from a controlled comparison.

Where the Ceiling Is and Why Pharma Is in No Hurry

Fully autonomous AI breakthroughs are still far off. On TCSExam, a benchmark of 100 theoretical computer science problems, leading models produced verifiable proofs, but did not solve a single open hypothesis on their own. They work confidently where the path is known, and so far do not go beyond the boundary of the known.

American competitors see similar limits. Anthropic's Claude found an enzyme system in September, but OpenAI's claims of breakthroughs in mathematics and physics drew criticism from mathematicians: there are individual impressive results, but the disputed claims are no fewer.

Pharma itself is not convinced either. In a June Citi survey of 50 heads of research units in the US, 72% already use AI widely, but only 24% expect it to noticeably raise the share of successful drug development efforts. AI is accepted as a working tool, but not as a guarantee that drugs will appear more often.

What This Means in Practice

First, don't count on model quality on coding benchmarks continuing to grow at the old pace: the developers themselves speak of a slowdown. Second, narrow scientific assistants like BioMaster look like a realistic niche: a well-defined domain, a measurable metric (hours of a specialist's work) and ready-made data. Third, such products should be sold as routine accelerators, not as "automated scientists": it is the speedup that has a verifiable effect, and the industry's skepticism is backed by the survey figures.

A caveat: most of the information comes from a company representative's talk and press retellings, and the BioMaster result is described by a single test without independent verification. Still, the direction is telling: the fight for LLM leadership is gradually shifting from the question "who writes code better" to the question "who is more useful in real science."

Story

Key Questions Answered

Why Moonshot Is Moving Away From Code?

Moonshot AI's international director, Giacomo Ficari, said at Web Summit in Lisbon that science requires smarter models than law, finance or programming. The logic: if an assistant for code or legal documents is already good enough for most tasks, further improvements deliver less and less noticeable benefit. Research work sets a higher bar, which leaves room to pull ahead of competitors. This is not an abandonment of code as such, but a change of priority.

What BioMaster Actually Showed?

It is important to understand what exactly got faster. The model did not make a discovery: it sifted a gene dataset and proposed a list of 100 candidates, which a biologist then verifies personally. That is the familiar role of an assistant: take over routine first-pass processing and shorten the path to a meaningful hypothesis. The gain, from four hours to 90 minutes, more than a twofold speedup, is nice, but it is a single test, not proof that a whole lab would work this way.

Where the Ceiling Is and Why Pharma Is in No Hurry?

Fully autonomous AI breakthroughs are still far off. On TCSExam, a benchmark of 100 theoretical computer science problems, leading models produced verifiable proofs, but did not solve a single open hypothesis on their own. They work confidently where the path is known, and so far do not go beyond the boundary of the known. American competitors see similar limits.

ZK
Zhemal Khamidun — CPO at Alpina Digital, product manager of AlpinaGPT. Building AI products since 2019. Hamidun News is a pet project that grew into a production aggregator with 6,000+ readers.
✦ Unified AI Workspace · Nomad

Test Frontier LLMs & Autonomous Agents in One Workspace

Run Claude Opus 5.5, GPT-6, Gemini 3.1, and DeepSeek side-by-side in Nomad AI (nomadnet.ai) — unified model routing, deep research, and multimodal workflows under one subscription.

Try Nomad AI →Launch Nomad Agent
What do you think?
Loading comments…