Within two years, Unisound's large model revenue share jumped from 5.5% to 99.9%, completing its main business pivot. Growth now hinges on capability reuse, reshaping its valuation logic.
The simplest way to judge whether a company has truly changed is not to listen to how it talks about strategy, but to look at where its revenue comes from. Yunzhi Technology (Unisound) has provided a very steep set of numbers.
In 2024, revenue related to large models accounted for 5.5% of product revenue.
In the first half of 2025, that figure was 42.4%.
In the first half of 2026, it was 99.9%.
In two years, the figure went from 5.5% to 99.9%.
During the same period, the company's revenue reached RMB 562 million, up 38.7% year over year.
This means that Unisound has not simply grown a large model business alongside its original speech recognition and chip businesses.
Rather, its core revenue itself has nearly completed a transition.
If the market's first impression of the company still focuses on speech recognition and AI chips, then the question truly worth discussing is no longer "How is Unisound's large model business growing?"
Instead, it is:
Is the market still using the framework of the previous generation of companies to understand a company that has already changed its revenue structure?
What Changes First Is Not the Label, but the Revenue
For most companies, "transformation" often first appears in strategic messaging.
Business names change, product packaging changes, but the revenue still comes from the old businesses.
This time, Unisound is different.
In the first half of 2026, revenue from intelligent enterprise solutions reached RMB 478 million, accounting for 85.1% of total revenue; Token business revenue approached RMB 30 million; and edge-side AI revenue approached RMB 53.8 million.
These three business segments now form the main body of today's revenue.
The speech recognition and AI chip businesses that were once the most recognizable are now difficult to identify as independent revenue segments in the financial report.
This shows that, for Unisound, large models are no longer a "second growth curve" but are increasingly becoming the new foundation of the business.
But what is even more worth studying is how it managed to complete such a rapid transition in just two years.
The answer is actually not mysterious.
Unisound did not start building a large model company from scratch.
It repackaged what it had accumulated over the past decade or more into large models.
In the medical field, the company has served more than 400 medical institutions; in the government affairs field, it has accumulated a large number of real business scenarios; its medical knowledge graph alone covers 10.18 million medical concepts and 5.33 million medical terms.
These customer relationships, industry know-how, and data assets may have only been project capabilities in the speech recognition era.
In the large model era, they have become the production resources for training, post-training, and agent deployment.
This is also why Unisound has a natural difference from many large model startups that started from scratch:
Others first have a model, then look for scenarios.
Unisound first had the scenarios, then put the model into them.
This path is not unfamiliar in the global market.
Palantir was long understood as a government project-based software vendor. Its true revaluation only came when AIP reorganized the customers and industry capabilities it had accumulated over many years into an AI platform.
It did not discard its historical baggage.
Instead, it turned its historical accumulation back into assets for the AI era.
Unisound is facing a similar issue today.
The business transition has already happened.
The market label has not yet fully caught up.
The More Important Change Is That the Growth Model Has Begun to Shift
Revenue coming from large models only proves that the transformation has occurred.
The more important question is:
Is this a better business?
In this interim report from Unisound, what is truly worth looking at is another set of numbers.
In the first half of the year, R&D expenses were RMB 284 million, up 69% year over year.
At the same time, revenue from existing customer renewals and upsells accounted for more than 60% of total revenue; sales and marketing expenses were RMB 41.91 million, growing at a slower pace than revenue; implementation costs fell from 21% to 17%; and the demand research cycle was shortened from 30 days to 15 days.
The company also disclosed that it has accumulated 1,773 instantiated agents, with the same module reused up to 119 times.
Put together, these numbers are more meaningful than looking at revenue growth alone.
They show that Unisound is shifting from a typical project-based growth model to one based on "capability reuse."
The growth model of traditional AI projects is simple:
One more customer means one more solution set, which means one more implementation team.
Revenue grows, and so do headcount and delivery costs.
In essence, it is a linear business.
But the real value that models and agents should create is to allow one set of capabilities to be used repeatedly.
Once a module is built, it should not just serve one hospital;
once a model is trained, it should not just serve one project.
R&D investment happens upfront, but revenue can keep coming from reuse later.
If this holds true, the company's operating model will change:
Revenue growth no longer requires sales and delivery costs to grow at the same rate.
This is one of the most important commercial implications of large models.
That is also why, even though Unisound significantly increased R&D investment in the first half of the year, gross profit growth has already outpaced revenue growth, and the loss rate has improved noticeably.
What is truly worth watching here is not any single financial metric.
It is a trend:
Technology investment is beginning to translate into operating efficiency.
This is a completely different story from the old logic of "burn more money to get more scale."
Three Companies Are Answering Three Different Commercialization Questions
If we broaden our view to the entire large model industry, this shift becomes even clearer.
In the first half of 2026, several large model companies have begun to deliver different commercialization answers.
Zhipu AI is closer to a path of directly monetizing model capabilities.
Through Tokens, APIs, and model services, it sells the technical capability itself. What it needs to prove is:
Can frontier models directly become a sufficiently large infrastructure business?
MiniMax is taking another path.
It quickly acquires global users through consumer-facing products, then keeps expanding into subscriptions and the enterprise market. What it needs to answer is:
Can a massive user base ultimately be converted into stable, high-quality revenue?
Unisound is closer to a third path.
Rather than extracting large models from industry scenarios, it embeds model capabilities into its existing high-value scenarios such as healthcare and government affairs, monetizing through agents, industry solutions, and Tokens together.
It proves:
Can the industry assets built over the past decade or more achieve higher reuse rates in the large model era?
There is no need to declare a winner among these three paths right now.
But they already show that the commercialization of large models in China is undergoing a clear divergence.
In the past, everyone was answering the same question:
Whose model is stronger?
Now they are answering different questions:
Who can sell the models?
Who can keep users?
Who can sell the same model capability over and over again?
This is also the sign that industry competition is truly becoming more complex.
Large Models Are Moving from an "Intelligence Race" to an "Efficiency Race"
Over the past three years, the most important variables in the large model industry have almost all been technical.
Parameter size, benchmarks, context length, reasoning capability, coding, agents.
The market first needed to confirm one thing:
Can Chinese companies build world-class large models?
But as more and more models enter a similar capability range, a new question will inevitably arise:
How does technical capability ultimately make its way into financial statements?
Because for a commercial company, benchmarks are always only an intermediate variable.
The true outcome variables are just a few:
Are customers willing to pay?
Will they keep buying after the first purchase?
Can one set of capabilities serve more customers?
Do costs rise at the same rate as revenue grows?
Do profit margins improve as scale expands?
In other words, in the next stage, what truly determines the value of a large model company may no longer be just "how strong the intelligence is."
Instead, it will be:
How efficiently can a unit of intelligence be produced, and how many times can it be sold again?
This is also why the significance of companies like Palantir goes beyond being just a valuation benchmark.
The more important lesson is:
Only when AI capabilities truly enter customer workflows and begin to improve revenue quality and operating efficiency will the market reassess the nature of a company.
The same applies to Chinese large model companies.
Zhipu AI needs to prove that models themselves can be scaled into revenue.
MiniMax needs to prove that users can be scaled into revenue.
Unisound needs to prove that industry accumulation can be scaled into reuse through models.
On the surface, three companies are taking three different paths.
Look deeper, and they are all answering the same question:
Besides becoming smarter, can large models actually make a company more efficient?
This may be the real watershed for the large model industry after 2026.
In the previous stage, the market was looking for the smartest models.
In the next stage, the market will increasingly look for:
Who can turn intelligence into a better business.