文 | 司库财经，作者 | 庆福，编辑 | 嘉嘉
2026年的国产大模型开始变得“不务正业”：10月13日，阶跃星辰与阿里千问在同一天发布了主打AI的手机和眼镜耳机产品。
两家大模型厂商在同一天发布电子消费产品绝非凑巧，而是代表着国内AI大模型的风向标——厂商之间的比拼将不只是最新版本AI大模型的性能参数，也不只是产品的价格成本，而是延伸到了消费硬件终端。
为什么在2026年，大模型厂商纷纷跨界下场做消费级的硬件产品？
原因也很简单，AI大模型更迭速度越来越快，这就导致各大厂商都面临一个问题，用户留不住，无法实现变现，因此抢占手机、眼镜等硬件入口，就成为必选题。
但是一堆软件应用公司扎堆做消费级的AI硬件，真的能将流失的用户闭环锁住吗？
高度竞争下的焦虑
在国内科技圈，AI大模型是最火的行业，没有之一，国家网信办公布的数据显示：
截至2026年7月，国内发布的大模型总数达1509个，位居全球第一；仅在2026年上半年，AI领域融资金额达4544.9亿元，几乎是去年的2倍。
数据上看，AI大模型似乎正处于繁花似锦、烈火烹油的黄金时代。
但正如《围城》一样，外人看到的都是美好，只有内行人才懂这些数据背后的辛酸——行业太卷了，卷到只要落后一步就会失去用户。
据《新智元》统计，从今年年初到9月22日，包括豆包、阿里千问、DEEPSEEK在内的十家主要厂商共发布了至少28个旗舰模型，平均每9.5天就有一个国产大模型发布。
而在价格上，2026年上半年，行业共发生六轮降价事件，其累计降幅超过2024年初至2025年底所有降价幅度的总和。
不难看出，性能和价格的内卷让整个大模型行业在不断发展，从宏观层面上是件好事，逼着各大国产大模型厂商提高竞争力。
但对于大模型厂商来说就未必了——因为这种迭代速度非常快的竞争，让厂商留不住用户，用户只会流向智能体验更好、成本费用更低的大模型产品。
QuestMobile的数据显示，今年1—2月，国内主流AI应用的首日留存率为65%，但是如果将时间拉长，其15日留存率只有18%，30天留存率只有12.8%。
要知道，一个产品要想赚钱，首先要有一定数量的基本盘用户，只有基本盘用户达到一定数量，企业才能够有想象空间，通过附加服务实现商业变现。
例如根据用户环境、行为和偏好数据，提供商品推荐或本地化吃喝玩乐推荐的服务；根据用户的语言习惯、工作风格提供办公助手等。
上述一切的畅想都是建立在，用户对某款大模型深度依赖的基础上，但显然在这个百花齐放的浪潮中，国内大模型厂商最缺的就是用户忠诚度和粘性。
押注消费级硬件入口
查漏就要补缺，怎么样才能解决用户留不住的问题呢？国内大模型厂商们将目光投向了消费硬件。
原因很简单，一是国内大模型更新迭代速度太快了，用户天然就会流向测试排名最高、性能最好的大模型应用，二是软件这种东西，用户更换的成本几乎为零。
而消费级AI硬件正好能够解决这两个问题，不论是AI眼镜还是搭载智能体的手机，用户都会在较长时间内持有并使用，这就导致，用户一旦使用自家的AI硬件产品，因为更换成本等原因，用户就很难流失。
“停留在App层的AI永远是‘租客’，只有自己做硬件，AI才能成为‘原住民’”，阶跃星辰CEO印奇的发言最直白，只有拿下消费级的AI硬件，大模型厂商才能实现用户的长久留存。
其实印奇的硬件理念早已经成为行业共识：
2025年12月，字节联合中兴努比亚推出搭载豆包手机助手的nubia M153，今年9月，豆包手机上市，京东预约量超36万台，并联合中国移动联合发布“灵犀·星火智盒”系列，硬件形态覆盖AI办公本、AI录音笔、AIPC等；
7月，阶跃星辰发布“大模型原生AI终端品牌”STEPX和其自研的首款手机 STEPX Neo，内置个人智能体Amoo。10月13日，STEPX Neo“先行版”正式发布，其全栈定价为3999元，Amoo可跨应用自主规划、调度对应服务，一站式完成订票、办公、视频剪辑、本地生活等多类任务；
10月，阿里发布日常“Personal Agent（个人智能体）”的千问AI眼镜N1/N1 Pro及首款千问AI耳夹式耳机。而据《晚点LatePost》报道，继AI眼镜后，阿里云旗下无影团队正在试水一款QwenBook的AI平板设备，定位原生智能体电脑，完整版预计2026年年底或2027年年初发布。
从字节豆包到阶跃星辰再到阿里千问，他们都在2026年上半年发布消费级的AI硬件，显然自研的消费级硬件已经成为国产AI大模型厂商的主要趋势。
理想面临新挑战
其实很多AI大模型厂商都将用户流失、商业难以变现等问题指向了消费级的AI硬件，试图通过终端硬件产品实现锁住用户，建立以自有大模型为核心的生态体系。
理想是很美好，但这张消费硬件反哺软件应用的大饼真的好画吗？我看未必。
首先是有限的资源限制，AI手机、AI眼镜等消费产品，要涉及供应链、研发、零配件备货等众多环节，这势必会牵扯大模型企业的人力和资金。
要知道，现在的大模型领域是“朝发夕至”的迭代速度，大模型厂商分散精力跨界做硬件，例如2025年，智谱营收只有7.24亿元，净亏损47.18亿元，亏损是营收的6.5倍；MiniMax总收入7903.8万美元，年内总亏损达18.72亿美元……
即便是像字节这样不差钱的大厂，豆包手机也不是全凭自己研发，是同中兴努比亚合作，对于阶跃星辰这类独立大模型公司来说，一边是持续的亏损，一边是消费硬件的高投入和高失败率，其独立做消费级AI硬件的压力可想而知。
其次除了大模型厂商在做AI硬件外，硬件厂商也在推出AI大模型，并将其植入自己的产品内部。
如小米MiMo系列，今年9月Artificial Analysis 智能指数中，小米 MiMo-V2.6-Pro获得46分，超越Kimi K3和GLM-5.3，位列全球开源模型第一、国产模型第一。
更重要的是，MiMo大模型已经被植入到小米各大硬件产品中，如小米18 Fold和小米18 Pro搭载了MiMo端侧大模型的超级小爱2.0；其首款AI原生手机“龙虾”已启动封测……
面对小米这类硬件能打，软件也不弱的选手，国产大模型厂商如何通过AI软件的优势吸引用户，取得优势才是未来行业关注的重点。
2026年的这场终端竞赛，表面上看是手机、眼镜、耳机的形态之争，实际上比的还是同一件事：谁能让用户真正离不开。 抢入口只是第一步，把入口变成习惯，把习惯变成生态，才是终局。
本文由 AI图鉴自动抓取并翻译，版权归原作者所有；如需查看原文请点下方链接。
Wen, the treasurer, the author, Qin Fong, the editor, Qajia.
The national production of LLM in 2026 began to become “irregular”: on October 13th, the skyrocketing star and Alian asked for the same day publish called the main AI phone and the headphone product.
The two LLM manufacturers did not coincide with the same day, but rather represented the windmark of the country's AZLLM - a comparison between manufacturers that would extend not only to the performance parameters of the latest version of AZLLM, but also to the price costs of the product, but to the end of the consumer hardware.
Why, in 2026, did LLM manufacturers cross borders to make consumer-grade hardware products?
The reason for this is also simple: AZLLM is changing at an increasingly rapid pace, which leads to a problem for all major manufacturers, where users cannot afford to stay in order to be able to cash, and thus the need to seize access to hardware such as mobile phones, glasses and so on.
But a bunch of software application companies stacked up to make consumer-grade AI hardware, can they really lock the lost users in the loop?
Anxiety in high competition
AZLLM is one of the most popular industries in the country's scientific community, and none of them has the data published by the National Network Office:
As of July 2026, the total number of LLMs in publish in the country was 1509, the first in the world; in the first half of 2026 alone, the amount of AI financing amounted to $45.449 billion, almost twice as much as last year.
According to the data, AZLLM seems to be in the golden age of flowering, burning and cooking.
But as in The Wall, outsiders see beauty, and only in-houses understand the bitterness behind these data — the industry is too thick to be able to lose its users if one step behind.
According to the NSW, from the beginning of this year until 22 September, 10 major manufacturers, including Doubao, Ali Millennium and DEEPSEK, had at least 28 flagship models of publish, an average of one national production of LLMpublish per 9.5 days.
In terms of prices, in the first half of 2026, there were six rounds of price reductions in the industry, with cumulative reductions exceeding the sum of all price reductions between the beginning of 2024 and the end of 2025.
It is not hard to see that the incisor of performance and price has allowed the entire LLM industry to develop and is good at the macro level, forcing the major producers of LLM to become more competitive.
But this is not necessarily true for LLM producers — because of the very rapid pace of this iterative competition, which makes it impossible for the manufacturers to keep the users, who will only go to AI agent, which is a better and less costly product.
Data from QuestMobile show that, between January and February this year, the first-day retention rate for mainstream AI applications in the country was 65 per cent, but if the time was extended, it would be 18 per cent for 15 days and 12.8 per cent for 30 days.
It is important to know that, in order for a product to make money, there must be a certain number of basic disk users, and only if the basic disk users reach a certain number, will the enterprise have the imagination to realize commercial liquidity through additional services.
For example, services based on user environment, behaviour and preferences are provided for recommended goods or locally recommended food and drink; and office assistants are provided according to user language habits, work styles, etc.
All of the above is based on the depth of user dependence on a certain LLM, but it is clear that, in this wave of flowers, the country ' s Z43Z manufacturer lacks the greatest user loyalty and viscosity.
Bet on the consumer hardware entrance.
How can we solve the problem of failure by filling the gaps? Domestic LLM manufacturers have turned their eyes to consumer hardware.
The reason for this is simple: first, the domestic LLM upgrades are too fast, and the users flow naturally to the highest-ranked and best-performing LLM applications, and second, software, the cost of which is almost zero.
Consumer-level AI hardware, on the other hand, can solve these two problems, and users will hold and use it for a longer period of time, whether it be AIS glasses or a mobile phone carrying AI agent, which makes it difficult for users to lose once they use their own AIS hardware products, for reasons such as replacement costs.
“AAI staying at the App level is always a “tenant” and only by making its own hardware can AI become an `indigenous people'”, the most straightforward statement by CEO Inch is that the long-term retention of the user can be achieved only by taking a consumer-grade AI hardware, the LLM manufacturer.
Indge’s hardware concept has long been agreed upon:
In December 2025, byte jointly launched Nubia M153, who carries Doubao mobile phone assistants. In September this year, Doubao mobile phones were put on the market, with over 360,000 reservations made in Kyoto, and the United China mobile joint publish series of “Linghingh-Silk Boxes”, with hardware forms covering AI office books, Ai audio pens, AIPC;
In July, STEPX, a step-by-step star, publish, “LLM Native AI Terminal Brand”, and its self-study first-hand mobile phone, STEPX Neo, incorporated an individual, AI agentA Amoo. On 13 October, STEPX Neo's “first-line” official publish, which is fully priced at US$ 3999, Amo can cross the application of autonomous planning, dispatch counterpart services, and all-in-one completes multiple types of assignments such as booking, office, video clipping, local life, etc.;
In October, the daily “Personal Agent (individual AI agent)” of Ali Z94Z asked about AI glasses N1/N1 Pro and the first one asked about AI headphones. However, according to LateLatePost, after the AI glasses, the Alibaba Cloud team is testing an AI tablet of QwenBook to locate the original AI agent computer, the full version of which is expected to be in late 2026 or early 2027.
From byteDoubao to Step Star, to Ali, they all spent the first half of 2026 on publish-grade AI hardware, apparently self-researched consumer hardware has become the dominant trend for the national production of AILLM manufacturers.
Ideals face new challenges
In fact, many AZLLM manufacturers point to consumer-grade AI hardware for loss of users, difficulty of commercial realization, and try to lock users through end-of-pipe hardware products, creating an ecological system centred on LLM.
Ideas are wonderful, but is this really good for a big pie that consumes the hardware re-energizing software? I don't think so.
First, there are limited resources, and consumer products such as AAI mobile phones, Ai glasses and so on involve the supply chain, R & D, spare parts, etc., which will involve the human and financial resources of LLM enterprises.
As you know, the area of LLM is now an iterative speed of “up to dawn”, with LLM producers dispersing their energy to cross-border hardware. For example, in 2025, Zhipu received only $724 million, resulting in a net loss of $4,718 million, or 6.5 times the amount received; Mini Max received $7,9038,000 in total, with a total loss of $1,872 million during the year.
Even large plants like byte, which do not develop their own mobile phones on their own, are working with independent LLM companies, such as China-Hinung Nubiya, with persistent losses on the one hand and high input and failure rates on the other.
Second, apart from the fact that the LLM manufacturer is working on AI hardware, the hardware manufacturer is introducing AILLM and embedding it within its own product.
Mi-V2.6-Pro received 46 points in the Artificial Analysis smart index in September of this year, exceeding Kimi K3 and GLM-53, ranking first in the Global Open Source Model and first in the National Production Model.
More importantly, MiMo LLM has been implanted in various millet hardware products, such as Mi18 Fold and Mi18 Pro, carrying the Super Love 2.0 of LLM on MiMo's end; its first AI primary phone, Lobster, has been activated...
The way in which domestic LLM manufacturers attract users through the advantages of AI is the focus of future industry attention in the face of hard-core and soft-soft hardware such as rice.
This end-of-pipe competition in 2026, which appears to be a struggle over the form of mobile phones, glasses, headphones, is actually the same thing: who can really keep the user apart. It is only a first step to turn the entrance into a habit and to turn it into an ecological one that is the end.
Auto-fetched and translated by AI Atlas. All rights belong to the original author.