UBS kept Buy ratings on MINIMAX-W and Z.AI even as the two Hong Kong-listed Chinese AI model developers fell about 20% and 40% month-to-date, with short-selling ratios of 9.652% and 11.761%. The bank attributed the drawdown to regulatory debate over open-source weighted models and tighter global anti-distillation measures, while pointing to overseas cloud partnerships and a pipeline of model launches as offsetting catalysts.
Two of Hong Kong's listed Chinese AI model developers have lost a combined slice of market value this month that no single regulatory headline can yet justify, and the positioning data shows traders are not waiting for one. Z.AI (智谱, 02513.HK) has shed roughly 40% month-to-date, while MINIMAX-W (00100.HK) is down about 20%, as a widening policy debate over open-source weighted models and tighter anti-distillation enforcement abroad reshapes how investors price Chinese capability catch-up.
UBS kept Buy ratings on both names, arguing that overseas expansion is converting open-source adoption into revenue. The bank flagged MiniMax's global enterprise adoption, vertical fine-tuning and cloud deployment partnerships, Z.AI's recent tie-ups with global cloud service providers, and Kimi K3's launch on Amazon Web Services' Bedrock platform as evidence that monetization is broadening beyond China.
The selling pressure is visible in positioning data. Short selling in Z.AI reached $970.28 million, equal to 11.761% of turnover, while MINIMAX-W short selling totaled $285.69 million, or 9.652%, according to short-selling data as of Sept. 23, 2026 at 16:25 HKT. Both ratios sit well above the levels typically seen in Hong Kong's large-cap technology complex, where short interest in the most heavily traded names rarely clears 5% of turnover.
What is at stake is whether the market is repricing a regulatory risk that has not yet become a rule, or a genuine slowdown in the pace at which Chinese labs close the gap with U.S. frontier models. UBS framed the outlook as two-sided: near-term pressure from policy uncertainty, offset by a pipeline of model launches the bank expects to repair investor sentiment.
The debate centers on whether open-source weighted models — the release format that let DeepSeek, Alibaba's Qwen and Z.AI's GLM family build global developer bases at near-zero distribution cost — will face restrictions. UBS said heightened discussion of such limits "adds uncertainty to overseas adoption," without specifying which jurisdiction or mechanism is under consideration.
Running alongside that is a tightening of anti-distillation measures globally, with U.S. policy paying greater attention to industrial-grade distillation. Distillation is the practice of training a smaller model on the outputs of a larger one, and it is the mechanism by which several Chinese models reached competitive benchmark scores at a fraction of frontier training cost. If enforcement tightens, the pace of capability catch-up becomes a function of domestic compute and data rather than access to frontier outputs.
That distinction matters for valuation. A model that catches up through distillation carries a shorter, cheaper development cycle; one that must close the gap through original pre-training requires far more compute, which is precisely the constraint U.S. export controls were designed to impose. UBS said recent fundraising by Chinese labs should help expand computing power and support training of next-generation larger-scale models — an implicit acknowledgment that the compute build-out is now the binding variable.
UBS's monetization case rests on distribution rather than model quality. Kimi K3's availability on AWS Bedrock puts a Chinese open-weight model inside the infrastructure that Western enterprises already use, removing the procurement friction that has historically kept Chinese models out of regulated corporate environments. Z.AI's cloud partnerships follow the same logic, and MiniMax's enterprise adoption and vertical fine-tuning work targets the segment where buyers pay for adaptation rather than raw inference.
The pricing structure of that business is deteriorating at one end and holding at the other. UBS expects per-token pricing at the scale end to remain deflationary, while pricing power at the capability frontier should prove more durable. In practice that means high-volume, low-complexity inference keeps getting cheaper — a margin problem for anyone selling commodity tokens — while the small number of tasks requiring frontier capability retains pricing power. For MINIMAX-W and Z.AI, the question is what share of their revenue mix sits at each end.
The launch calendar is the near-term swing factor. UBS named DeepSeek V4.1 Pro, Kimi K3.1, MiniMax M3 Pro and Z.AI GLM-5.5/6.0 as potential sentiment catalysts. Model releases have historically produced sharp single-day moves in Chinese AI names, and a strong GLM-5.5 or M3 Pro print would give the bulls a benchmark argument that does not depend on regulatory clarity.
For now, the tape is doing the talking. Z.AI's 40% drawdown and MINIMAX-W's 20% decline have both outpaced the broader Hang Seng Tech Index over the same period, which tells you the selling is name-specific rather than a sector rotation. With short interest at 11.761% and 9.652% respectively, any positive regulatory signal or strong model launch would force covering — but the same positioning means further headlines on open-source restrictions will be met with additional supply. Investors should watch for any formal policy language on open-source weights, and the next model releases from the UBS catalyst list, as the two events most likely to break the current range.
This article is for informational purposes only and does not constitute investment advice.
