Chinese large-model vendors shift from competing on models to grabbing consumer hardware entry points

In 2026, domestic large-model makers began to become "distracted from their main business": on October 13, StepFun and Alibaba's Qwen released AI-focused phones and glasses/earbuds products on the same day.

It is no coincidence that two large-model vendors released consumer electronics products on the same day; rather, it represents a bellwether for domestic AI large models—the competition among vendors will no longer be limited to the performance parameters of the latest AI large-model versions, nor to product price and cost, but will extend to consumer hardware terminals.

Why, in 2026, are large-model vendors crossing sectors one after another to make consumer-grade hardware products?

[1] Anxiety under intense competition

As of July 2026, the total number of large models released in China reached 1,509, ranking first in the world; in the first half of 2026 alone, financing in the AI field reached 454.49 billion yuan, almost twice that of last year.

But just like in Fortress Besieged, outsiders see only the beauty, and only insiders understand the bitterness behind these figures—the industry is too cutthroat, so cutthroat that falling behind by even one step means losing users.

According to statistics from AI Era, from the beginning of this year to September 22, ten major vendors including Doubao, Alibaba Qwen, and DEEPSEEK released at least 28 flagship models in total, with a domestic large model released on average every 9.5 days.

(Source: AI Era)

In terms of price, in the first half of 2026, the industry saw six rounds of price cuts, with a cumulative reduction exceeding the sum of all price cuts from early 2024 to the end of 2025.

It is not hard to see that the involution in performance and price has kept the entire large-model industry developing, which is a good thing at the macro level, forcing major domestic large-model vendors to improve their competitiveness.

But for large-model vendors, this is not necessarily the case—because this extremely fast-iterating competition makes it hard for vendors to retain users, and users will only flow to large-model products with better intelligent experience and lower costs.

QuestMobile data shows that from January to February this year, the first-day retention rate of mainstream domestic AI applications was 65%, but if the time span is extended, the 15-day retention rate was only 18%, and the 30-day retention rate was only 12.8%.

You should know that for a product to make money, it first needs a certain number of core users; only when core users reach a certain number can a company have room for imagination and achieve commercial monetization through additional services.

For example, based on user environment, behavior, and preference data, it can provide product recommendations or localized recommendations for food, drink, and entertainment; based on users' language habits and work styles, it can provide office assistants, etc.

All the above visions are built on the basis that users are deeply dependent on a certain large model, but obviously, in this wave of blooming flowers, what domestic large-model vendors lack most is user loyalty and stickiness.

[2] Betting on consumer-grade hardware entry points

If there are gaps, they must be filled. So how can the problem of failing to retain users be solved? Domestic large-model vendors have set their sights on consumer hardware.

The reason is simple: first, domestic large models iterate too quickly, so users naturally gravitate to the large-model applications with the highest test rankings and the best performance; second, with software, the cost for users to switch is almost zero.

Consumer-grade AI hardware happens to solve both problems. Whether it is AI glasses or phones equipped with intelligent agents, users will hold and use them for a relatively long time. As a result, once users use a company's own AI hardware products, they are hard to lose because of switching costs and other reasons.

"AI that stays at the App layer is always a 'tenant'; only by making hardware ourselves can AI become an 'indigenous resident,'" said Yin Qi, CEO of StepFun, most bluntly. Only by securing consumer-grade AI hardware can large-model vendors achieve long-term user retention.

In fact, Yin Qi's hardware philosophy has long become an industry consensus:

In December 2025, ByteDance teamed up with ZTE nubia to launch the nubia M153 equipped with the Doubao phone assistant. In September this year, the Doubao phone went on sale, with more than 360,000 reservations on JD.com, and it also jointly released the "Lingxi·Xinghuo Smart Box" series with China Mobile, with hardware forms covering AI office notebooks, AI voice recorders, AI PCs, and more;

In July, StepFun released "the large-model-native AI terminal brand" STEPX and its self-developed first phone, STEPX Neo, with the built-in personal agent Amoo. On October 13, the "Pioneer Edition" of STEPX Neo was officially released, with a full-stack price of 3,999 yuan. Amoo can autonomously plan across applications and dispatch corresponding services, completing tasks such as ticket booking, office work, video editing, and local life services in a one-stop manner;

(Source: StepFun official website)

In October, Alibaba released the Qwen AI glasses N1/N1 Pro, which daily "Personal Agent" and its first Qwen AI clip-on earphones. And according to a report by LatePost, after AI glasses, Alibaba Cloud's Wuying team is testing the waters with a QwenBook AI tablet device, positioned as a native intelligent agent computer, with the full version expected to be released at the end of 2026 or early 2027.

[3] Li Auto Faces New Challenges

The ideal is beautiful, but is this big pie of consumer hardware feeding back into software applications really easy to draw? I don't think so.

You should know that the large-model field now iterates at a "depart at dawn and arrive at dusk" speed, and large-model vendors are dispersing their energy to make hardware across sectors. For example, in 2025, Zhipu's revenue was only 724 million yuan, with a net loss of 4.718 billion yuan, 6.5 times its revenue; MiniMax's total revenue was 79.038 million US dollars, with a total loss of 1.872 billion US dollars during the year...

Even for a wealthy big company like ByteDance, the Doubao phone was not entirely developed on its own; it was in cooperation with ZTE nubia. For independent large-model companies like StepFun, with continuous losses on one side and the high investment and high failure rate of consumer hardware on the other, the pressure of independently making consumer-grade AI hardware is imaginable.

Take Xiaomi's MiMo series. In the Artificial Analysis Intelligence Index in September this year, Xiaomi MiMo-V2.6-Pro scored 46 points, surpassing Kimi K3 and GLM-5.3, ranking first among global open-source models and first among domestic models.

(Source: Xiaomi MiMo official website)

More importantly, the MiMo large model has been embedded into Xiaomi's major hardware products, such as the Xiaomi 18 Fold and Xiaomi 18 Pro, which are equipped with Super Xiao Ai 2.0 powered by the MiMo on-device large model; its first AI-native phone, "Lobster," has started closed beta testing...

The terminal competition in 2026 appears on the surface to be a battle over the forms of phones, glasses, and earphones, but in reality it is still comparing the same thing: who can make users truly unable to do without them. Grabbing the entry point is only the first step; turning the entry point into a habit and the habit into an ecosystem is the endgame.

This article comes from the WeChat public account "Siku Finance," author: Qingfu, editor: Jiajia, published by 36Kr with authorization.