36Kr just broke the news that Unisound's U2 is coming, with an official launch in June, described as the pivotal move to break into the domestic first tier. Based on this account, U2's technical capabilities are benchmarked against the level of top-tier large models.
U2's model performance has reached the level of world-class large models, with comprehensive performance benchmarked against top global models across key evaluations including instruction following, knowledge reasoning, and Agent tool invocation. More importantly, U2 is a native Agent-capability model that can enter business workflows, understand objectives, break down tasks, invoke tools, and complete execution and validation, making it closer to a "foundation model for the agent era." Moreover, compared to trillion-parameter models, its scale is far smaller, with parameter efficiency nearly five times higher.
U2 is Unisound's DeepSeek moment, driven by the same logic: smaller parameters, lower costs, but higher value per Token.
This is the core of density intelligence: not simply emphasizing a smaller model, but achieving world-class capability with smaller scale and lower cost.
If the news is accurate, the market now faces an interesting pricing anomaly: a model company about to officially enter the first tier is trading at a price-to-sales ratio of under 15 times.
This is not a signal that can be casually ignored.
The Pricing Anomaly
Zhipu AI and MiniMax have both recently listed on the Hong Kong Stock Exchange, with valuations exceeding HKD 100 billion. Baichuan and Moonshot AI are still racing toward IPOs, with valuations also at the hundred-billion level. Unisound, meanwhile, has a price-to-sales ratio of under 15 times. This valuation gap is plainly visible.
The financial reports tell a different story. In 2025, Unisound's full-year large model revenue reached RMB 610 million, accounting for 50.4% of total revenue, up 1076% year-over-year. That growth rate is rare among any Hong Kong-listed tech company. And this RMB 610 million was not subsidized by financing. Full-year adjusted net profit continued to improve, with an 84% improvement in the second half.
But the market does not seem to have caught on yet.
Some might say Hong Kong-listed companies are conservative in pricing AI stocks. True, but conservatism and neglect are two different things. A leading model company in its industry, with large model revenue up tenfold and rapidly improving margins, yet its market cap stuck at this level, looks more like an information gap. Most people in the market have not yet equated Unisound with the first tier of large model companies.
The Hong Kong discount is not entirely a valuation system issue. Unisound's narrative has long been tied to labels like "speech recognition" and "intelligent voice." Most people still think of it as a speech chip company; the large model story has not reached the capital markets.
The company filed for listing in 2021 but only officially listed in June 2025—a four-year IPO marathon that happened to miss the most active publicity window for large models. During that period, Zhipu secured major funding from Zhongguancun and Alibaba, and Moonshot AI received USD 1 billion from Alibaba. Unisound has consistently kept a low profile, and compliance requirements further narrowed its window for external communication.
Unisound founder and CEO Huang Wei said in an interview with LatePost: "2023 to 2025 was the warm-up for large models; the real game only begins in 2026." During the warm-up, Unisound made little external noise, avoided funding battles, and simply went hospital by hospital, refined scenario by scenario, and signed contract after contract. It was not until the 2025 financial report came out that people realized large model revenue accounted for half of total revenue, up 1076% year-over-year. When the market is flooded with grand narratives, the company quietly getting things done gets overlooked.
The official launch of U2 will be the starting point for closing this information gap.
The Moat in Large Model Deployment
The most troublesome issue with large models is hallucination—making things up. In a medical context, a fabricated answer can be a matter of life and death.
In 2025, over 450,000 medical records at Beijing Friendship Hospital were generated by Unisound's AI, with a direct physician adoption rate exceeding 90%. In auto insurance claims, cost control rates improved by 3 percentage points compared to traditional third-party management companies. In medical insurance review, cost control rates rose from 1 percentage point to 7 percentage points.
These numbers are backed by real money. All of that RMB 610 million in large model revenue comes from high-value scenarios: medical record generation, auto claim assessment, and medical insurance review. In these scenarios, a few dozen characters of output can trigger decisions involving hundreds of thousands of yuan, which is a completely different value proposition from conversational tokens.
So why Unisound? Can't open-source large models also get the job done?
The issue is that medical AI relies on more than just language understanding—it requires a deep grasp of medical workflows, clinical standards, and insurance policies. General models are trained on public internet data and have never experienced real hospital working environments. Unisound has深耕 the medical and IoT sectors for thirteen years, accumulating actual usage data from over a thousand hospitals. This data includes not just text, but also physician usage habits, hospital workflows, and the specific needs of different departments. The moat formed by this deep integration is unlikely to be replicated by open-source large models.
The arrival of U2 makes this moat even deeper. Its hallucination resistance, supported by thirteen years of scenario data, is the foundation of its industry leadership in medical and financial scenarios.
The Growth Engine
U2 also has a compelling cost story. The 36Kr report mentioned that its per-Token cost can be reduced to about one-tenth that of a typical dense model. Translate that into real enterprise usage, and the financial significance becomes clear.
Unisound's sales expense ratio is only 5.4%, nearly the lowest among B2B AI companies, where peers typically run above 20%. Per-capita output is RMB 2.52 million, R&D staff account for 69% of the workforce, and R&D investment represents 75% of total adjusted three-item expenses. Taken together, the logic is simple: after U2 goes live, Unisound's scale expansion no longer depends on throwing more people and money at the problem.
Growth is driven by the competitiveness of the product itself. According to sources close to the company, Unisound's growth in Token revenue is even more aggressive than outside estimates. In May 2026, the company's ARR from Token call revenue surged 600% month-over-month, and June forecasts are even more aggressive, with month-over-month growth still doubling.
Behind this growth is U2's rollout pushing Unisound from a project-based model to full productization. Huang Wei mentioned in the LatePost interview that a medical AI project used to take three months; now, with U2's standardized capabilities, delivery cycles have been shortened to one week, and costs reduced by 80%. U2 is not just a technology upgrade—it directly rewrites the company's delivery logic.
Based on the nearly USD 5 million ARR in May, the current market cap corresponds to a Token price-to-sales ratio of under 20 times. By comparison, first-tier peers typically trade above 50 times. Unisound's Token revenue growth is on par with any first-tier player, yet the capital market's pricing falls far short.
The Window at the Inflection Point
Full-year 2025 total revenue was RMB 1.21 billion, with large model-related revenue at RMB 610 million, up 1076% year-over-year. Customers are renewing contracts, and new customer wins are expanding.
Full-year adjusted net profit continued to improve. The improvement was concentrated in the second half: the first half faced margin pressure due to IPO expenses and concentrated R&D investment in large models. After U2 launched in the second half, revenue accelerated, standardized delivery cut project costs by 80%, and second-half profit improvement reached 84% year-over-year. This pace is approaching breakeven. Once profitability is achieved, Hong Kong valuation logic shifts from PS to PE, leaving considerable valuation upside.
Huang Wei has said an AI company's value equals intelligence density multiplied by Token value. That is where U2's confidence comes from: model performance has reached world-class large model standards, yet the parameter count is just 260 billion—far smaller than those trillion-parameter models—and Token costs are reduced to one-tenth that of dense models. First establish a world-class capability benchmark, then talk about smaller scale and lower cost—that is the premise on which density intelligence can hold. Strong yet compact: with equivalent performance, parameter scale is smaller, with parameter efficiency nearly five times higher; strong yet economical: per-Token cost drops to one-tenth that of dense models, making it genuinely affordable for enterprises; strong yet deployable: native Agent capabilities let the model truly enter business workflows, understand objectives, break down tasks, and invoke tools, rather than staying confined to a chat window. Combined, these three dimensions complete the full logic of density intelligence. The Tokens produced directly serve high-value business scenarios: medical records, insurance claims, and medical cost control—Token value in these scenarios is on an entirely different level from conversational tokens.
He also once said: you don't need an academician of the Chinese Academy of Sciences to drive a ride-hailing car. Large models don't need to be universally capable; achieving excellence in specific scenarios delivers greater commercial value. Unisound has pursued this path for 13 years, and U2 is the most important step along it.
Solutions revenue accounts for 69.8%, and product revenue for 10.1%. A high solutions share means strong customer stickiness and significant switching costs—once integrated, future revenue is locked in. Add in the visibility boost from U2's official launch and the warming sentiment in Hong Kong's AI sector, and there is no shortage of drivers for valuation recovery.
Final Thoughts
The profitability inflection point is accelerating. Margin improvement was already very fast in the second half, and if growth continues, the flexibility will materialize the moment profitability is realized, with valuation logic shifting from PS to PE.
For Unisound, this is not a sentiment-driven rally. The industry is shifting from "looking at parameters" to "looking at deployment," and Unisound has been doing deployment for 13 years—the data from medical cost control, auto claim assessment, and medical record generation speaks for itself.
A company with large model revenue up tenfold, continuously improving margins, and U2 about to launch officially—none of these changes are reflected in its pricing. When this valuation gap closes depends on three things: when the profitability inflection arrives, whether U2 can sustain the growth acceleration, and when the market will catch on.
The market has offered a discount, but that discount does not account for these changes. What remains is simply when the market will turn around. But at least Unisound has done its homework—U2 is the answer sheet.
