Beijing-based ModelBest spent August in pre-IPO tutoring with China's securities regulator, positioning itself as the first large-language-model developer to test Shanghai's new STAR Market channel for unprofitable technology firms 3 6. Three weeks later, on September 9, it did something an ordinary IPO candidate would not: it gave away the crown jewel. The company and its OpenBMB research partner released MiniCPM5-2B, a 2-billion-parameter edge model, as a fully open-sourced pipeline — training-data curation, pre-training and alignment all public — pushing cumulative downloads of the MiniCPM family past 50 million 1. Six weeks earlier that count stood at 38 million 4. The stock is not public yet. The model already is.

A chip strategy built on everyone else's silicon

ModelBest's pitch is not that it owns compute, the way Nvidia or SMIC do. It is that MiniCPM runs efficiently on compute it does not own. The company has adapted its models to Qualcomm, MediaTek, Intel, Rockchip, Nvidia and AMD chip platforms, and in July it confirmed that MiniCPM would ship inside Samsung's Galaxy AI stack — a rare instance of a Chinese foundation-model maker embedded directly in a flagship non-Chinese consumer device, running alongside Qualcomm's expanded Snapdragon deal with Samsung across phones, watches and smart glasses 2 5. That is the commercial logic behind the giveaway: MiniCPM does not need to win a model-quality leaderboard, it needs to become the default small model that hardware makers reach for when they want on-device inference without sending user data to a cloud server. Downloads are the funnel; licensing and systems-integration deals with device makers are meant to be the monetization at the other end. At July's World AI Conference in Shanghai, ModelBest extended that funnel into robotics, unveiling the MiniCPM-Robot embodied-AI series aimed at the same edge-hardware logic — cheap, local inference — applied to machines instead of phones.

Why Shanghai just opened the door

The giveaway only makes sense against a regulatory shift most investors outside China have not priced in. On June 18, the Shanghai Stock Exchange activated STAR Market rules letting unprofitable AI and quantum-computing firms list, provided they meet market-capitalization and R&D-spending thresholds instead of a profit test 6. That is a direct answer to the fact that China's most valuable AI assets — foundation models, chip design, edge inference stacks — mostly do not generate GAAP profit yet, and were previously funneled to Hong Kong or Nasdaq to raise capital, as Zhipu AI and MiniMax have already done 3. ModelBest, which traces back to a Tsinghua University NLP lab and formally launched in 2022, is the test case regulators and bankers are watching to see whether a pure-play, largely open-source model company can clear the bar 3. Its most recent funding round, backed in part by China Telecom, pushed its valuation past $2.8 billion after the Galaxy AI news broke 5 2.

The revenue question the prospectus will have to answer

An open-source model with 50 million downloads is a strong distribution story and a weak line item. ModelBest's actual revenue comes from a narrower set of relationships: enterprise and automotive licensing deals, device-maker integration contracts like the Samsung and Qualcomm tie-ups, and government-linked capital that treats the company as strategic infrastructure rather than a pure financial bet — its investor roster includes state-linked funds alongside China Telecom 5. That is a materially different revenue mix from a subscription-API business like OpenAI's or Anthropic's, and it means ModelBest's STAR Market prospectus will need to show growth in paid integration and hardware-royalty revenue that is currently invisible in any public download counter. Competition is not waiting for that disclosure. Alibaba pushed a laptop-optimized, fully open-weight Qwen release in August explicitly aimed at the same on-device battleground, backed by a balance sheet ModelBest cannot match 7. Xiaomi, MiniMax and Zhipu are all shipping competitive small models into the same phone-and-PC channel, and unlike ModelBest, several already carry public listings and disclosed revenue that analysts can underwrite 3.

What investors should watch

Three things will determine whether ModelBest's STAR Market push succeeds where a pure download count cannot. First, the prospectus itself: China's exchange will require disclosure of R&D spending and a credible path to the market-cap and revenue thresholds the new rule sets for unprofitable issuers, which will be the first hard numbers the public gets on how much of the Samsung and China Telecom relationships actually convert to cash 6. Second, whether rivals with stronger balance sheets — Alibaba most obviously — can commoditize small on-device models faster than ModelBest can convert downloads into paid hardware deals, compressing the licensing economics the whole open-source strategy depends on 7. Third, precedent risk and reward: if ModelBest lists successfully, it becomes the template other loss-making Chinese AI labs use to reach public markets at home instead of Hong Kong, widening the investable set; if its numbers disappoint, regulators have a ready excuse to slow-walk the next unprofitable-tech applicant. For a market structurally short of ways to buy exposure to China's edge-AI buildout, ModelBest's listing is likely to be treated as a bellwether well beyond its own market cap.