NeuroAI NEUROAINEUROAI.SITE
China AI, reported clearly
1,300+ verified stories on models, robotics, BCI and the business of AI.
Browse stories
ESC

A 56-billion-parameter model built to design one enzyme at a time

Tianjin University released Tianhua Kaiwu, a 56-billion-parameter enzyme design model that produces a design in a week and a strain in three months.

2026-11-20 · 4339 words · NeuroAI
A 56-billion-parameter model built to design one enzyme at a time

On November 19, Tianjin University unveiled a model called Tianhua Kaiwu, with 56 billion parameters.

Tianjin Daily, Science and Technology Daily and China News Service all used the same phrasing: the first publicly released enzyme-design large model in synthetic biology.

Enzymes are the working tools of industrial biology. Laundry detergents need enzymes that survive heat; pharmaceutical intermediates need enzymes with extreme stereoselectivity; materials production needs enzymes that tolerate organic solvents. The traditional route is to search nature, then test candidates one by one.

From 800 million to 56 billion parameters

According to Science and Technology Daily, the university began the work in 2021, produced an 800-million-parameter base model in 2024, reached 7 billion in 2025, and then expanded to 56 billion with Zhipu AI.

Seventy times in three years.

That curve is unremarkable for a general-purpose model. It is worth a second look in a vertical domain, because the training data was not scraped.

Yuan Yingjin, an academician of the Chinese Academy of Sciences who directs the university's national key laboratory for synthetic biology, said the model was built from the architecture up rather than fine-tuned from an existing one.

The training base is enzyme sequences, structures, reactions and process data accumulated in laboratories and on production lines. That corpus is orders of magnitude smaller than web text, and general models cannot reach it.

One week for a design, three months for a strain

The published timeline: a design proposal within a week, and customized strain construction plus process development within three months.

The model is wired to the university's automated experimental platform, which closes three loops. It designs new enzymes from a target reaction and predicts mutation effects. It converts digital designs into physical samples through automated gene synthesis, assembly and expression. And it feeds experimental results back into training, so accuracy compounds.

There is an arithmetic detail that is easy to miss. Fifty-six billion parameters serve "dozens of enzymes" — the figure Yuan gave for tests completed so far.

Fifty-six billion divided by a few dozen leaves more than a billion parameters per enzyme.

That division is mine, based on public numbers, and it is not the university's framing. But it says something: this is a base model plus task heads, not one model swallowing the field.

What gets compressed is design, not validation

Between the one-week design and the three-month strain sits wet-lab work.

The model can say which sequences might work. The lab still has to synthesize them, express them, and measure activity. Throughput at that step sets the real speed of the system.

Narrowing candidates from thousands to dozens does not shorten the queue if the bench capacity stays the same.

At the launch, one company representative argued that once on-demand design stabilizes, buyers will move from "buy the enzyme, then rework the process" to "set the spec, then order the enzyme." If that holds, it changes procurement, not just R&D.

Representatives from PetroChina, China Merchants Group, Shanghai Pudong Development Bank, WuXi AppTec, Novo Nordisk and Asymchem attended. The guest list is itself a signal: the customer for enzymes is a production line, not a lab bench.

Domestic compute and private deployment

The model runs on domestic accelerators and supports on-premise deployment, so customer data never leaves the plant.

For biomanufacturing firms, that clause may matter more than the parameter count. Strain and process data are core assets.

Next comes open trials with research and industry partners, plus joint work on enzyme libraries and application standards.

Standards are harder than models. A model can be retrained in a year; a standard takes years of negotiation.

Why this data cannot be scraped

General models learn from text that already exists. Enzyme design learns from measurements that have to be made.

A sequence, a structure, a reaction condition and a yield are four different kinds of record, often produced by different instruments in different years. Assembling them into a training set is the expensive part, and the resulting corpus belongs to whoever ran the experiments.

That is the moat the university describes. It is also the limit: a model trained on one lab history learns that lab blind spots along with its knowledge.

The three-month timeline for strain construction depends as much on the automation platform as on the model. Liquid-handling robots, automated gene synthesis and expression screening set the pace. Without that hardware, a one-week design is a document, not a product.

The collaboration model is worth noting too. Beijing develops the model, Tianjin converts it, industry applies it, a division of labour that mirrors how biomanufacturing capacity is distributed.

*Disclaimer: This article draws on the Tianjin University launch and reporting by Tianjin Daily, Science and Technology Daily and China News Service. It is for reference only and does not constitute investment advice.*

Related coverage

Cite this article

NeuroAI (脑机网), "A 56-billion-parameter model built to design one enzyme at a time," 20 November 2026, https://neuroai.site/a/na-app-enzyme-design-1120

Machine-readable: BibTeX · RIS · JSON · Markdown

More in “AI in Action” → · Back to home · Markdown version

China AI, reported clearly
1,300+ verified stories on models, robotics, BCI and the business of AI.
Browse stories