Has a new player joined the first echelon of large models, and the market hasn't reacted yet?

According to the news just reported by 36Kr, Unisound's U2 is coming and will be officially launched in June. It is described as a crucial battle for Unisound to enter the first - tier of domestic models. By this account, the technical capabilities of U2 are on par with those of the first - tier large models.

The model performance of U2 has reached the level of world - class large models. In multiple key evaluations such as instruction following, knowledge reasoning, and Agent tool invocation, it has comprehensive performance comparable to that of the world's top models. More importantly, U2 is a native Agent - capable model. It can enter business processes, understand goals, break down tasks, invoke tools, and complete execution and acceptance. It is closer to the "base model in the era of intelligent agents". Moreover, compared with those models with trillions of parameters, its scale is much smaller, and the parameter efficiency is nearly 5 times higher.

U2 is Unisound's DeepSeek moment. The same logic applies: fewer parameters, lower costs, but higher value per Token.

This is the confidence of density intelligence: it's not simply about emphasizing a small - scale model, but achieving world - class capabilities with a smaller scale and lower cost.

If the news is true, the market is currently facing an interesting pricing misalignment: a model company on the verge of officially entering the first - tier has a price - to - sales ratio of less than 15 times.

This is not a signal that can be easily ignored.

Pricing Misalignment

Zhipu AI and MiniMax have successively listed on the Hong Kong Stock Exchange, with valuations exceeding HK$100 billion. Baichuan and Yuezhianmian are still sprinting for IPOs, and their valuations are also in the tens of billions. However, Unisound's price - to - sales ratio is less than 15 times. This price difference is obvious.

The financial report tells another story. In 2025, Unisound's annual large - model revenue was 610 million yuan, accounting for 50.4% of the total revenue, a year - on - year increase of 1076%. Such a growth rate is rare among any Hong Kong - listed technology companies. Moreover, this 610 million yuan was not obtained through financing subsidies. The annual adjusted net profit continued to improve, with an improvement rate of 84% in the second half of the year.

However, the market doesn't seem to have reacted yet.

Some may say that the Hong Kong stock market has always been conservative in pricing AI companies. That's true, but conservatism is different from neglect. For an industry - leading model company with a ten - fold increase in large - model revenue and a rapid improvement in profitability, but with its market value remaining stagnant, it seems more like an information gap issue. Most people in the market haven't equated Unisound with the first - tier of large - model companies.

The discount in the Hong Kong stock market is not entirely due to the valuation system. Unisound has long been labeled as a "voice recognition" and "intelligent voice" company. Most people still think it is mainly engaged in voice chips, and the news about its large - model business hasn't reached the capital market.

The company submitted its IPO application in 2021 and was officially listed in June 2025, a four - year IPO journey that happened to miss the most active dissemination window for large - model news. During this period, Zhipu received large - scale investments from Zhongguancun and Alibaba, and Yuezhianmian received US$1 billion from Alibaba. Unisound has always kept a low profile, and compliance requirements have further narrowed its scope of external communication.

Huang Wei, the founder and CEO of Unisound, said in an interview with LatePost, "From 2023 to 2025 was the warm - up stage for large models, and the real competition started in 2026." During the warm - up stage, Unisound didn't make much noise externally or engage in a financing war. Instead, it visited hospitals one by one, refined each scenario, and signed contracts one by one. It wasn't until the 2025 financial report was released that people realized that large - model revenue accounted for half of the total revenue, with a year - on - year increase of 1076%. While the market was bombarded with various grand narratives, the company that focused on doing things was ignored.

The official release of U2 will be the starting point for bridging this information gap.

Barriers to Large - Model Implementation

The most headache - inducing aspect of large models is their tendency to generate nonsense. In the medical scenario, a nonsense statement could be a matter of life and death.

In 2025, more than 450,000 medical records in Beijing Friendship Hospital were generated by Unisound's AI, and the direct citation rate by doctors exceeded 90%. In the auto insurance claims scenario, the cost - control rate increased by 3 percentage points compared with traditional third - party management companies. In the medical insurance scenario, the cost - control rate increased from 1% to 7%.

There is real money behind these numbers. Every bit of the 610 million yuan in large - model revenue comes from high - value scenarios: medical record generation, auto insurance damage assessment, and medical insurance review. In these scenarios, an output of just a few dozen words could trigger a decision - making amount of hundreds of thousands of yuan, which is completely different from the commercial value of chat - style Tokens.

Then why Unisound? Can't open - source large models do the job?

The problem is that medical AI relies not only on language understanding ability but also on a thorough understanding of medical processes, clinical norms, and insurance policies. General models learn from publicly available data on the entire network and have no experience in actual hospital work scenarios. Unisound has been deeply involved in the medical and Internet of Things fields for 13 years and has accumulated actual usage data from thousands of hospitals. This data includes not only text but also doctors' usage habits, hospital work processes, and special requirements of different departments. The barriers formed by this in - depth integration are unlikely to be replicated by open - source large models.

The arrival of U2 has made this barrier even higher. The anti - hallucination ability supported by 13 years of scenario data is the foundation for its leading position in the medical and financial scenarios.

Growth Engine

U2 also has an edge in terms of cost. The 36Kr report mentioned that its unit Token cost can be reduced to about one - tenth of that of ordinary dense models. When this figure is applied to the actual corporate usage scenario, the significance of cost savings becomes apparent.

Unisound's sales expense ratio is only 5.4%, which is almost the lowest among B - end AI companies. The ratio of its peers is usually over 20%. The per - capita output value is 2.52 million yuan, with R & D personnel accounting for 69% and R & D investment accounting for 75% of the total adjusted three - item expenses. Overall, the logic is clear: After the launch of U2, Unisound's scale expansion no longer needs to rely on increasing the number of people and investment.

Growth depends on the competitiveness of the product itself. According to insiders close to the company, Unisound's growth rate in Token revenue is even more impressive than the outside world's speculation. In May 2026, the ARR of the company's Token invocation revenue skyrocketed by 600% month - on - month, and the forecast data for June is even more aggressive, with a month - on - month doubling growth.

Behind this growth rate is the fact that the launch of U2 has promoted Unisound's full - scale transformation from project - based to product - based. Huang Wei mentioned in the LatePost interview that previously, it took 3 months to complete a medical AI project. Now, with U2's standardization capabilities, the delivery cycle has been shortened to 1 week, and the cost has been reduced by 80%. U2 is not just a technological upgrade; it has completely rewritten the company's delivery logic.

If calculated based on the ARR of nearly $5 million in May, the price - to - sales ratio of Token revenue corresponding to the current market value is less than 20 times. However, the peers in the first - tier generally have a ratio of over 50 times. Unisound's growth rate in Token revenue is comparable to that of any first - tier company, but the pricing given by the capital market is far from sufficient.

Inflection - Point Window

In 2025, the total annual revenue was 1.21 billion yuan, and the large - model - related revenue was 610 million yuan, a year - on - year increase of 1076%. Customers are continuing to renew their contracts, and the acquisition of new customers is also on the rise.

The annual adjusted net profit continued to improve. The improvement was concentrated in the second half of the year. In the first half of the year, due to IPO expenses and concentrated investment in large - model R & D, the profit margin was under pressure. After the launch of U2 in the second half of the year, revenue accelerated, and the standardized delivery reduced project costs by 80%. The profit improvement rate in the second half of the year reached 84% year - on - year. This speed is approaching the break - even point. Once profitable, the pricing logic of the Hong Kong stock market will shift from PS to PE, and the valuation flexibility will be significant.

Huang Wei said that the value of an AI company is equal to the intelligent density multiplied by the Token value. This is where U2's confidence lies: its model performance has reached the level of world - class large models, but its parameter scale is only 260 billion, much smaller than those models with trillions of parameters, and the Token cost has been reduced to one - tenth of that of dense models. First, establish the benchmark of world - class capabilities, and then talk about a smaller scale and lower cost. This is the prerequisite for density intelligence to be established. Strong and small: with the same performance, the parameter scale is smaller, and the parameter efficiency is nearly 5 times higher; strong and cost - effective: the unit Token cost is reduced to one - tenth of that of dense models, making it truly affordable for enterprises; strong and implementable: the native Agent ability enables the model to truly enter business processes, understand goals, break down tasks, and invoke tools, rather than staying in the dialog box. Only when these three aspects are combined can the complete logic of density intelligence be realized. The generated Tokens are directly applied to high - value commercial scenarios: medical records, insurance claims, and medical insurance cost control. The value of Tokens in these scenarios is on a completely different level from that of chat - style Tokens.

He also said, "You don't need an academician from the Chinese Academy of Sciences to drive a taxi." Large models don't need to be all - around and general; achieving excellence in specific scenarios can bring greater commercial value. Unisound has been following this path for 13 years, and U2 is the most important step on this journey.

Solution revenue accounts for 69.8%, and product revenue accounts for 10.1%. A high proportion of solution revenue means strong customer stickiness and high replacement costs. Once a customer is acquired, subsequent revenue is locked. Coupled with the increased attention brought by the official release of U2 and the warming - up of the AI sector in the Hong Kong stock market, there is no shortage of motivation for valuation repair.

Conclusion

The profit inflection point is approaching rapidly. The profit improvement rate was very high in the second half of the year. If the growth rate continues, the valuation logic will shift from PS to PE once the profitability is realized.

For Unisound, this is not just a short - term market sentiment. The industry is shifting from "focusing on parameters" to "focusing on implementation". Unisound has been working on implementation for 13 years, and the data from scenarios such as medical cost control, auto insurance damage assessment, and medical record generation speak for themselves.

For a company with a ten - fold increase in large - model revenue, continuous improvement in profitability, and the upcoming official release of U2, these changes are hardly reflected in the pricing. When this valuation gap will be filled depends on three things: when the profit inflection point will come, whether U2 can continue to boost the growth rate, and when the market will react.

The market has given a discount, but these changes are not factored into it. It remains to be seen when the market will turn around. At least, Unisound has done its part, and U2 is the answer.

This article is from the WeChat official account "Business Can't Be Cold", written by Liu Yike and reprinted by 36Kr with permission.