---
title: "MiniMax lists in Hong Kong one day after Zhipu — and calls the model the product"
date: 2026-11-15
category: Foundation Models
site: NeuroAI
canonical: https://neuroai.site/a/na-model-minimax-hk-ipo-models
language: en
---

# MiniMax lists in Hong Kong one day after Zhipu — and calls the model the product

> Shanghai's MiniMax (稀宇科技) debuted in Hong Kong at HK$165 on a full stack of text, video, speech and music models — not one app.

A 385-person company whose employees average 29 years old walked onto the Hong Kong exchange floor and priced itself above HK$100 billion in a single trading day. The business behind the ticker 0100.HK is MiniMax (稀宇科技), and almost none of its value comes from a product you can point to on an app store.

It comes from the models themselves.

## A listing measured in days, not decades

MiniMax listed on January 9, 2026, one day after Zhipu, at an offer price of HK$165 (≈ US$21.20) per share — the top of its range. With the greenshoe fully exercised, the deal raised about HK$5.54 billion (≈ US$710 million). Shares more than doubled on the first day, closing at HK$345 (≈ US$44.20) for a market value near HK$106.7 billion (≈ US$13.7 billion), a gain of roughly 109%.

Founded in early 2022, MiniMax reached a public listing in about four years, a pace its own filings and local press described as the fastest ever for an AI model company. The founder, Yan Junjie (闫俊杰), is a former SenseTime vice president. The anchor investors read like a map of China's internet: Alibaba, Tencent, and MiHoYo among the cornerstone buyers, with GIC, Baillie Gifford, and Norway's central bank also taking part.

## The portfolio, not the app

Where Zhipu led with one flagship large model (大模型), MiniMax sells a stack.

- **M-series text models** — the M2 family and its 2026 updates (M2.5, M2.7, and the M3 line) for general reasoning and coding.

- **Hailuo (海螺) video** — a text-to-video model the company says is among the few that can faithfully simulate gymnastics motion.

- **Speech (语音)** — a text-to-speech model MiniMax says has led global leaderboards ahead of OpenAI and ElevenLabs.

- **Music** — an audio generation model for short-form content.

MiniMax has also published open-weight (开源权重) versions of earlier text models, a habit that matters in a market where "download and self-host" is a selling point for buyers who will not send data to a third party. The company's stated strategy is "model-as-product" (模型即产品): the model is the unit shipped, wrapped in thin consumer apps like Talkie and Hailuo AI rather than the other way around.

The text side has moved quickly. MiniMax shipped the M2 family in early 2025 and, across 2026, rolled out the M2.5, M2.7, and M3 updates, each tightening reasoning and coding while preserving the MoE "small model, big capability" economics that let a 2.4-trillion-class result compete on price. The video and speech arms are the differentiator: Hailuo is pitched at filmmakers and advertisers, while the Speech model targets dubbing and assistive reading.

## How big, and how global

The numbers MiniMax disclosed point to a company already operating at scale:

- Roughly 212 million individual users and 130,000 enterprise clients across more than 200 countries and regions.

- Revenue in the first nine months of 2025 grew more than 170% year on year, with overseas sales above 70% of the total.

- Its open platform processes more than one trillion tokens a day.

Those figures explain the valuation better than any single demo. A model company that earns in dollars from users in dozens of countries looks less like a science project and more like infrastructure.

## The parts the prospectus cannot hide

The same filing that supports the story also lists the risks. MiniMax posted a net loss of about US$512 million in the first nine months of 2025, with research spending running at more than three times revenue — normal for a frontier lab, but a real burn. On the legal side, the company faces a copyright suit in the United States filed in 2025 by Disney, Universal, and Warner Bros. over generated video, and a separate 2026 accusation from Anthropic that it used fraudulent accounts to distill competitor models. Both are unresolved and both bear directly on whether a model company can monetize content generation without inheriting the liabilities of the content industry.

According to Li An, chief scientist at BrainNet (脑机网), the MiniMax case is less about one model's quality and more about whether a publicly listed company can keep a multi-modal model stack funded while litigation and compute costs compound.

## Honest limitations

This article relies on MiniMax's listing documents, government and exchange-adjacent reporting (Xuhui district government, China Economic Net, Xinhua's Economic Information Daily), and encyclopedic summaries; it does not independently verify user, revenue, or token-volume figures, which are company-reported. The "fastest AI IPO" and "first large-model stock" labels depend on definitions — Zhipu and MiniMax listed within a day of each other, and both claim firsts on different grounds. Benchmark claims for Hailuo and Speech come from MiniMax's own marketing and were not reproduced by an independent lab in the sources reviewed. The legal cases are allegations and lawsuits, not findings of liability. The first-day share price and market cap reflect a single trading day and are not a stable valuation.

## What readers can do now

- **Builders:** try the M-series and Hailuo models through MiniMax's open platform and compare token cost and video fidelity against open-weight (开源权重) alternatives before committing.

- **Investors:** read the prospectus's loss and litigation lines as the real thesis — the question is whether multi-modal model stacks reach profit before the legal bills do.

- **Content teams:** if you plan to ship Hailuo-generated video at scale, get a copyright review first; the pending US suit is exactly the risk you would be inheriting.

---

Published by NeuroAI (https://neuroai.site/) — https://neuroai.site/a/na-model-minimax-hk-ipo-models
Free to quote with attribution and a link to the original.
