ModelBest Just Filed for an IPO It Is Not Profitable Enough For. Its AI Already Ships Inside a Samsung Galaxy.
By Michele De Filippo
06 Sep 2026

The filing

On August 12, 2026, Beijing-based ModelBest submitted IPO tutoring paperwork to the Beijing Securities Regulatory Bureau, starting the formal countdown toward a mainland listing 1 3. CITIC Securities is sponsoring the process, which followed an August 5 shareholding restructuring that cleared the last internal step before the filing 4. The mandatory tutoring period runs at least three months, which points toward an actual listing sometime in 2027 rather than an imminent debut 2.

That timeline detail matters less than the company's financial profile. ModelBest is, by its own filing, not yet turning a profit at meaningful scale. It is attempting this listing anyway because Shanghai's STAR Market maintains a special standard for high-quality technology companies that have not yet generated significant revenue, and ModelBest is shaping up as the clearest test yet of whether that carve-out works for a foundation-model company rather than a biotech or chip-equipment name 2 4. A clean listing would tell every other loss-making Chinese AI lab that Shanghai, not only Hong Kong, is now genuinely open for their capital raising. A stumble would say the opposite just as loudly.

The money behind it

The scale backing this bet is real. ModelBest raised more than RMB 5 billion (roughly USD 740 million) across multiple rounds in the first half of 2026 alone, pushing its valuation past RMB 20 billion (about USD 2.8 billion) and making it the highest-valued startup in China's on-device AI segment 3 5 6. Backers span state-linked capital tied to telecom and financial groups, regional venture funds, and, in a later round, an unnamed automaker, a mix that signals both policy favor and genuine commercial appetite for the underlying technology 5.

That technology is a deliberate bet against the industry's dominant instinct to scale up. ModelBest, incubated inside Tsinghua University's natural-language-processing laboratory and formally founded in August 2022, builds small models designed to run entirely on the device in front of a user rather than in a data center 3. Its MiniCPM family has been downloaded more than 38 million times across GitHub and Hugging Face, and the flagship MiniCPM-V variant packs roughly 1.3 billion parameters into a footprint light enough to run on a phone with just 6 gigabytes of memory, across iOS, Android and HarmonyOS 3 6.

Cleared for takeoff, literally, at the regulator

Edge deployment in China is no longer just an efficiency choice, it is increasingly the compliance-cleared path to shipping AI features on a domestic phone at all. On July 15, 2026, the Cyberspace Administration of China approved the first batch of on-device generative AI services for commercial filing, covering seven device makers: Apple, Huawei, Xiaomi, OPPO, vivo, Samsung and Nubia 7. Every model that wants to sit inside one of those devices now needs to clear that gate, and ModelBest has built its entire product line around exactly that kind of on-device deployment.

The Samsung wrinkle

The detail worth sitting with longest is Samsung. Independent reporting from both South China Morning Post and Pandaily describes ModelBest as a partner to Samsung and Huawei, with MiniCPM models integrated into upcoming Galaxy-line flagship devices, expanding Samsung's on-device AI sourcing beyond its existing Google relationship 2 6. ModelBest has also disclosed chip-level partnerships with Qualcomm and Intel, and product deployments stretching into automotive and education hardware well beyond smartphones 2.

Set against Washington's continuing push to keep advanced compute and, increasingly, model software out of what it calls adversarial hands, a South Korean device maker choosing a Chinese small-model vendor for on-device inference is a data point worth more than the deal's disclosed value, which neither side has published. It suggests the frontier-model export-control fight and the on-device small-model market are decoupling from each other. Model weights small enough to run entirely on a handset, and never touch a cloud server, are drawing a different level of scrutiny than the cloud-scale systems currently subject to sanctions and chip export limits.

What this signals for investors

Three threads converge in a single filing. First, Chinese capital markets are actively re-underwriting AI companies that do not yet carry the revenue profile public investors normally demand, and ModelBest's three-month tutoring window is the cleanest near-term test of whether that underwriting actually holds through to a completed listing. Second, the commercial center of gravity in consumer AI is visibly shifting toward the edge, where a model's ability to run on a mid-range phone chipset within a 6-gigabyte memory budget matters more than its benchmark score against the largest cloud systems. Third, the assumption that geopolitical friction cleanly separates Chinese and non-Chinese AI supply chains looks thinner with each such deal; a flagship Korean phone shipping a Chinese-built model is exactly the kind of cross-border technology dependency that trade policymakers in both Washington and Seoul will eventually have to price in.

For portfolios positioned around Asia's AI buildout, the ticker to watch over the next several quarters is less ModelBest itself than the STAR Market listing standard it is testing. If a pre-revenue foundation-model company can successfully float in Shanghai, expect a queue of similarly unprofitable Chinese AI startups, in embodied AI, in enterprise tooling, in other edge-model niches, to follow ModelBest's path rather than default to Hong Kong or New York, reshaping where the next generation of Asian AI capital actually gets raised.

Follow signals beyond the surface.
Learn how Midas turns market change into intelligence.