On April 14, 2026, Unisound's share price surged over 35% in a single day, hitting a two-month high. This capital market move was not an isolated event: its financial report showed a 92% narrowing of net losses, a more than tenfold surge in large-model revenue that crossed the 50% revenue contribution threshold for the first time, and the official launch of the desktop-level agent U2Claw on April 8—multiple signals released in rapid succession. The close alignment in timing of the technology release, financial data, and share price movement has been read by the market as a positive signal, but whether the short-term gains are sustainable will still need to be validated by earnings performance.
Meanwhile, market attention is shifting from model parameters to financial performance. This transition is especially evident in Hong Kong's "Big Three" large-model companies: Zhipu AI, MiniMax, and Unisound. All three went public on the Hong Kong Stock Exchange in succession starting in June 2025, each charting a distinct path to commercialization. Unlike other players still mired in computing cost and customer acquisition challenges, how is Unisound leveraging its "strong foundational model plus deep applications" strategy to move toward a platform-based model? This is not just a question for investors—it is a core issue as the large-model industry enters deeper waters.
The Big Three Diverge
Hong Kong currently lists three large-model companies: Zhipu AI, MiniMax, and Unisound. All three tell an AI story, but their paths to profitability are completely different. Their listing timelines also differ: Unisound went public first in June 2025, while Zhipu AI and MiniMax followed in January 2026.
Zhipu AI's Tsinghua University pedigree gives it full-stack self-developed capabilities. Its GLM model series iterates rapidly, with the goal of benchmarking against Silicon Valley. On the commercial front, Zhipu relies on its MaaS platform and on-premise deployments to secure custom orders from large financial and government clients. The advantage of this business is strong customer stickiness—once a system is integrated, switching vendors is difficult. But the downside is heavy costs: in 2024, Zhipu's computing service fees reached RMB 1.55 billion, accounting for 70.7% of R&D spending, and in the first half of 2025 that ratio stayed at 71.8%. Spending 70% of R&D funds on computing procurement is a capital-intensive, high-risk operating strategy.
Unlike Zhipu's asset-heavy model, MiniMax chose a different route: avoiding the heavy-delivery B2B business and focusing instead on consumer-facing applications combining text, video, and audio. Leveraging its early overseas operational experience, its products have attracted a large user base internationally, with overseas revenue now accounting for over 70% of total. This is a classic internet traffic model, but maintaining it comes at a high price—user acquisition spending and inference computing costs remain elevated, and consumer users are highly prone to churn, forcing the team to continuously invest in marketing and feature updates to retain users.
Unlike the aggressive approaches of the other two, Unisound is extremely pragmatic. Built on its "Shanhai·Atlas" general-purpose intelligent computing foundation, Unisound has not scattered its efforts but has focused on high-frequency, essential needs in healthcare, daily life, and transportation. Through its agent platform delivery model, it has evolved from early single-point custom projects to scaled services. The key to this path lies in the continuous accumulation of business data and improvements in operational efficiency. One core metric stands out: Unisound's per-capita output has risen to RMB 2.52 million per person. In an AI sector accustomed to burning cash, this operational efficiency could be one of its differentiating advantages—but whether this advantage can be sustainably converted into market valuation still depends on future earnings delivery.
Unisound's Financial Turning Point
To understand why capital suddenly "got" Unisound in 2026, one needs to look at the key changes in its 2025 annual report: the RMB 610 million in large-model revenue cited in that report signals that its technology has begun translating into scalable, replicable products—a breakthrough its earlier single-point project model failed to achieve.
The "Shanhai·Atlas" foundation demonstrated its business fit during the year. Unlike early models that remained on benchmark leaderboards, "Shanhai·Zhiyi 5.0" has entered real-world application scenarios. It integrates medical knowledge enhancement technology and ranked first in three categories—medical agent, medical large language model, and medical multimodal large model—in the MedBench 4.0 evaluation. The medical agent scored 94.6 points, passing clinical decision support tests in a laboratory setting—though its stability in real hospital environments still requires validation through more cases. In the voice domain, "Shanhai·Zhiyin 2.0" has compressed end-to-end latency to within 90 milliseconds while supporting over 30 dialects and 14 international languages. U1-OCR solved cross-industry document recognition at the 3B parameter scale. Individually, each of these technical metrics may not be eye-catching, but strung together across different scenarios, they already constitute a set of production-ready tools.
Building on this foundational capability, Unisound launched its agent platform in 2025, standardizing its business capabilities into packaged offerings. Its Claw platform, in essence, wraps complex models into directly callable modules. It has now integrated 450 hospitals, 110 million AI chips, and transportation dispatch systems across more than 10 cities. Engineers no longer need to hand-code at each site—the platform's APIs handle technology delivery and billing directly.
This business model is directly reflected in the financials. In the second half of 2025, its net loss narrowed 84% year-over-year, with adjusted losses narrowing 92%. During the same period, Zhipu's computing spending was still expanding, and MiniMax's overseas customer acquisition costs remained high. Unisound's losses were narrowing while the other two had not yet reached that inflection point. Amid a shared period of heavy computing investment, Unisound was the first of the three to achieve loss narrowing, setting its financial performance apart from peers. Breaking down the cost structure, Unisound's expense ratio dropped significantly by 10 percentage points, and selling expenses fell rather than rose, compressed to 5.4% of revenue. By generating stable recurring revenue through API calls and token billing—replacing field sales teams that require continuous investment—this asset-light, product-focused model is seen by some investors as a sign that the MaaS business is beginning to show economies of scale.
Technology Upgrades and Agents
In 2026, the center of gravity in the large-model industry is shifting. All three companies are advancing technology upgrades, but with different priorities: Zhipu plans to launch GLM-5, MiniMax is pushing M2.1, and Unisound is preparing to release Foundation 2.0 around June, followed by a native agent built on that foundation between the end of Q2 and Q3, targeting office and coding scenarios. The industry's competitive focus may be moving from parameter scale to capability density and inference cost.
Unisound's Foundation 2.0 is optimized for reasoning, coding, and multimodal capabilities—three areas pointing at a single goal: delivering more accurate computation and better answers at the same server cost. This approach directly targets a common industry pain point: the computing bill determines per-call cost, which in turn compresses or expands room for commercial pricing.
The core need in office and coding scenarios is not information retrieval but task completion. Unlike early Q&A assistants, native agents must execute processes autonomously: understanding ambiguous intent, planning paths, calling tools, and producing results. To control the cost of high-intensity autonomous operation, Unisound adopted context-aware compression technology, cutting token consumption in half, while adding data security mechanisms to meet enterprise private deployment requirements.
Huang Wei previously noted: "Vendors that cannot adapt to the agent ecosystem in the future will find it hard to survive." All three companies are currently building agents, but with different delivery formats. As a new interaction paradigm, agents are reshaping human-machine collaboration and also enabling Unisound to extend into consumer-facing scenarios such as office productivity.
From Vertical Depth to Ecosystem Building
In 2026, Unisound's commercialization map is expanding in multiple directions. On the B2B side, standardized modules for medical insurance and transportation are being replicated nationwide—provincial government projects have high entry barriers, but cracking one can generate a wave of follow-on business. On the consumer side, the U2Claw desktop agent launched on April 8 targets everyday office work with one-click installation and on-device data processing, marking Unisound's experiment in finding new growth. Overseas, it is advancing through the ASEAN Cooperation Center, with Zhiyin 2.0's multilingual capability serving as the entry point into Southeast Asian markets. On the open-source front, Skills has been integrated into ClawHub, with its enterprise-grade platform positioned around security and control, though the commercialization path is still being explored. With multiple lines running in parallel, whether it can establish a foothold in any single domain will depend on how orders materialize in 2026.
Final Thoughts
The three companies face different situations. Zhipu concentrates 70% of its R&D budget on computing procurement, but its technology moat has yet to generate scale returns, and losses remain under pressure. MiniMax's revenue is heavily dependent on overseas markets, leaving its business stability exposed to regulatory tightening. By contrast, Unisound has deepened its focus on verticals such as healthcare and transportation, using project accumulation to drive platform-based replication, and narrowed its losses by 92% in 2025—the first of the three to show signs of financial improvement.
The competitive focus of the large-model race is shifting from model capability showdowns to sustained validation of commercialization paths. Whether scale expansion can drive self-sustaining profitability will be the determining factor in which companies maintain a competitive edge during the current phase of industry consolidation.