In the best era of large model industrialization, Chinese AI has "killed" the worship of parameters.

In 2025, it is regarded as the year of "disenchantment" for large models.

The once-feverish imagination of AGI has been quickly replaced by industrial challenges such as "inference cost" and "practicality". From Silicon Valley to Beijing, Shanghai, and Guangzhou, no one is willing to easily pay for the simple "parameter competition" anymore. Large models are accelerating from simple technological exploration into the deep waters of commercialization where technology and demand are racing against each other.

This transformation has made giants like OpenAI and Google change their strategies and start to intensively launch cost - effective inference models for the enterprise - level market. The late - comer advantage of the Gemini series of models also proves to us that the main stage for the iteration of large - model capabilities may not be in the laboratory. In the past, it was like looking for nails with a hammer. Now, whether it's Gemini and the data empire Google behind it, or Alibaba Cloud and the Qianwen large model, which have been in the industrial digital - intelligence field for many years, they are all shaping the "hammer" according to the shape of the "nail".

Market adaptability has extremely compressed the model iteration cycle from several years to months or even weeks. And this shift in perspective has precisely created the best opportunity for Chinese AI to "overtake on a different track".

For a long time, the spotlight of global artificial intelligence has been on Silicon Valley. However, in 2025, a watershed has emerged: Chinese large models are no longer satisfied with copying and following technological paths. Instead, in the new cycle of shifting from parameter - centric to industry - centric, they have moved from passivity to proactivity. The computing power blockade has not defeated domestic large models; instead, it has brought unique resilience, specifically manifested as: pragmatism in architectural innovation, in - depth full - stack services in commercialization, and strategies and global responsibilities in ecological layout.

This long - term path makes the "Chinese solution" more likely to take the lead in this era of AI industrialization than ever before.

01. In the Deep Waters of Industrialization, Large Models are Becoming "Super Supporting Actors"

Why do large models need to be "disenchanted" in implementation?

Since 2025, an interesting phenomenon has emerged in the deep waters of industrialization: Artificial intelligence seems to be everywhere, but compared with the once - popular Chatbot that was discussed on every street corner, large models now seem to be "invisible" in terms of product form.

Behind this invisibility is the pragmatism of Chinese enterprises in industrial iteration and the full - stack and collaborative technical support of AI manufacturers.

As AI moves from the digital world to the physical world, the automobile is at the forefront. The "invisibility" trend of large models is most vividly reflected in the field of intelligent driving. Today's cars have gone beyond being just means of transportation and have become intelligent agents capable of active understanding, decision - making, and action. In this process, large models step back and become the "super foundation" driving the evolution of intelligent driving.

In the past year, whether traditional or new - energy vehicles, Chinese automakers have basically been competing in one thing: the implementation of intelligent features. With the support of manufacturers like Alibaba Cloud, Chinese automakers have achieved astonishing speed in the industrialization of large models in this imaginative physical space.

Take XPeng Motors as an example. The efficiency of intelligent iteration is a major technological advantage. XPeng's second - generation VLA (Visual - Language - Action Model) is trained on 100 million clips of data, equivalent to the sum of extreme scenarios that a human driver would encounter in 65,000 years of driving.

The first difficulty in intelligentization lies in the large - scale data throughput. With the support of Alibaba Cloud's cloud - based computing power cluster, XPeng has built the first AI computing power cluster of 10 EFLOPS scale in the domestic automotive industry, creating a "cloud - based model factory". The full - link iteration cycle from the cloud to the end can reach an average of once every five days. In XPeng's technological core, the significance of large models is not only the front - end experience and interaction but also the silent contribution in the operation of data and scenarios.

In the invisible cloud, relying on its full - stack AI cloud service capabilities, Alibaba Cloud has supported the intelligent implementation of 100% of Chinese automakers. While leading the Chinese driving R & D computing power market, globally, this industrialization achievement and speed are also far beyond the reach of other AI giants.

In the field of intelligent cockpits, which is closer to ordinary users, large models are also hidden under the interaction system of end - cloud collaboration and ecological integration, bringing more delicate changes to the relationship between humans and vehicles from a tool - based to a partner - based one. Large models are becoming the brain for the intelligent upgrade of the Chinese automotive industry in a more subtle way.

Turning our attention to the manufacturing industry, the influence of large models has extended from office efficiency to production lines. In the past year, Sany Heavy Industry has built more than 50 AI Agents based on the Qianwen large model, covering the entire chain of R & D, production, sales, and services. For industrial giants, the most valuable asset is the non - standard experience accumulated over decades. Through full - parameter fine - tuning, Sany Heavy Industry has "embedded" this industry know - how into the large model, internalizing the production data and expert experience accumulated by the enterprise over decades into replicable digital assets. From then on, the large model can guide novice engineers in equipment fault diagnosis and process optimization like a "veteran", enabling non - standard scenarios to be scaled up, but the real heroes are still the experience accumulated by enterprises and people in practice.

Industrialization is redefining technology, which is an indisputable fact. Therefore, in the deep waters of industrialization, large models must be willing to be supporting actors, a practical, useful, and affordable "super supporting actor".

The low latency required by factory assembly lines, the zero - tolerance for content hallucinations in financial risk control systems, and the high sensitivity of intelligent hardware to power consumption and complex environments in edge - side inference... To solve these problems, large models need to learn to move from burning money to implementation, from the grand narrative of general intelligence to the detailed fine - tuning of scenarios.

In 2025, the domestic large models that are most practical have best interpreted the word "implementation" in the global AI industry.

02. The Evolution of Large - Model Efficiency Determines Future Productivity

Looking at the history of technological development, the implementation of most new technologies is top - down. There are not many industries like large models that firmly link the first step of commercialization with "affordability".

The difference between this and consumer brands' rejection of premium is that the popularization of AI is not a market competition strategy but the goal of technological development.

In 2025, domestic large models redefined SOTA: The highest score on the list only shows that the technology is "useful", but to make the most useful technology accessible and affordable, the efficiency and productivity per unit of computing power are more worthy of attention as hard standards.

On the industrial side, business owners are not concerned about how many points a large model gets in a niche knowledge quiz but how much actual efficiency improvement they can get for each dollar invested. This is why domestic large models gave up technological show - off early on.

However, giving up technological show - off does not mean giving up on technology, and high cost - effectiveness does not mean the model is backward. On the contrary, if the technology and performance are not hard - core enough, industrial implementation can only be a castle in the air.

With the emphasis on computing power ROI, Chinese AI has gone further on the path of "anti - parameter competition" and achieved a dimensionality - reduction blow on the path of "efficiency competition".

In the past two years, the competition was about who had a smarter model, while in 2025, the competition is about who has a more cost - effective and faster model. A series of technological breakthroughs by Alibaba Cloud and the Qianwen large model are essentially aimed at solving the three major industrial obstacles of "high computing power cost, slow inference, and difficult deployment" for the industrial circle.

Data and performance are the most intuitive evidence. In the field of large models, technological show - off is often manifested as deliberately magnifying the contingency of getting high scores on a fixed benchmark test paper, but it has little practicality in the industrial field. The Qianwen large model chose an unpredictable real - combat environment. In the blind - test platform Chatbot Arena where real users randomly set questions, it ranked among the top three globally, won first place in five capabilities such as code and mathematics, and set a new record for domestic models, breaking through the siege of closed - source models as an open - source model.

When the mainstream voice in the industry was still focused on parameter competition and anxious about training models with hundreds of billions of parameters, Qianwen 3 began to innovate in architecture through a hybrid inference architecture and high - sparsity MoE technology. By compressing the intelligence of 80 billion parameters into an inference process that only needs to activate 3 billion parameters each time, the training and inference costs were reduced by 90%. This achievement enables Qianwen 3 - Next, which is targeted at small and medium - sized enterprises with low budgets, to match the efficiency of the Qianwen 3 - 235B flagship model.

In terms of multi - modal controllability, the industrial difficulties are more prominent. The high - value - density scenarios of multi - modal generation are in industrial fields such as film and television production and human - machine interaction, but its "black - hole - like" computing power consumption, inefficiency, and uncontrollability seriously hinder the realization of value. In the past year, as a representative of domestic "full - scale, full - modal" models, the Qianwen large model has clearly focused its iteration on overcoming industrial difficulties. Wanxiang Wan 2.2 has become the industry's first MoE video - generation model, directly saving 50% of computing power; Wanxiang Wan 2.6 has become the first domestic video model supporting role - playing functions, showing excellent control over the physical world and overcoming industrial difficulties such as audio - video synchronization and multi - shot generation; Fun - Audio - Chat - 8B has achieved SOTA in multiple professional benchmark tests of audio models.

The efficiency evolution of computing power is also reflected in the further blurring of the boundary between the cloud and the edge. Last year, the release of models such as Qianwen 3 - 4B and Qianwen 3 - 30B - A3B enabled domestic large models with performance comparable to GPT - 4o to run smoothly on consumer - grade hardware such as mobile phones. For developers, this means releasing computing power from the expensive APIs of cloud servers and replacing it with efficient and low - cost local capabilities.

In addition to being "cost - effective", there is also "speed". In long - text processing, Qianwen 2.5 - 1M has increased the inference speed by nearly 7 times, changing the experience of processing million - word - level contract documents from "minutes" to "seconds". The improvement in speed is of decisive value in time - sensitive businesses such as financial risk control and data compliance.

The ultimate manifestation of computing power efficiency is the qualitative change in productivity.

Therefore, the SOTA results are just the appearance, and the story of productivity improvement behind them is the true nature of Chinese AI. The future of computing power efficiency evolution will be manifested as a qualitative change in productivity. For example, in the field of AIGC, after Haiyi AI connected to Wanxiang 2.6, users' content creation efficiency increased by 6 times; in the field of human resources, Zhaopin used AI to create multiple intelligent agents, increasing the average matching rate of jobs and candidates by more than 70%.

Compared with Silicon Valley, which is still anxious about whether the Scaling Law has reached its peak, the current characteristic of Chinese AI is that it has opened a new growth curve through the "engineering dividend". The Best Paper Award at NeurIPS won by Alibaba Cloud in 2025 and the SOTA achieved in global authoritative evaluations are also the natural results of its "pragmatism" approach.

Because in the context of industrialization, the best technology is always the technology that can be most effectively transformed into advanced productivity.

03. The Open - Source Ecosystem Enables Chinese AI to Adopt a Long - Term Approach

Technology determines how high a manufacturer can go, while the ecosystem determines how far it can go.

In addition to the returns from architectural innovation and industrial delivery capabilities, the ecosystem is a key structural keyword that cannot be ignored for the late - comer advantage of domestic large models. For example, while global giants are turning to closed - source models, Alibaba Cloud has persisted in developing a series of open - source large models.

Open - source means actively giving up the window period of code leadership, which seems to be a "hell - mode" from the outside world. However, in Alibaba Cloud's strategic plan, open - source is a very effective ecological competition strategy. This is why, as a player in the first echelon of domestic large models, the Qianwen large model has maintained its first - mover advantage in technology and achieved leading market shares in multiple industries.

When evaluating the competitiveness of an AI enterprise, such as answering why it is difficult for AI startups to compete with "big players", the first things we pay attention to are capital reserves and resource centralization. However, behind money and resources, there are three key hidden conditions: a more comprehensive technology supply chain, a more active ecosystem, and more stable industrial - level standards.

Alibaba Cloud is currently the only "big player" in the world that simultaneously has full - modal open - source large models and full - stack AI services. The ecological position of having models on one hand and infrastructure on the other gives the Qianwen large model unique competitiveness in the underlying logic of computing power scheduling and application development. The positioning of a full - stack AI service provider also gives Alibaba Cloud a moat of "experience and courage" in this new cycle: it is not afraid of being copied and caught up in technological innovation, and not afraid of losing orders due to open - source - based independent development in commercial implementation.

The rewards for this courage were reaped in 2025. Thanks to open - source, the Qianwen large - model series has been tested and modified by global developers in various extreme scenarios, undergoing an ecological - level "crowdsourced iteration". This has enabled Qianwen to evolve faster than closed - source models and systems in the laboratory, and even promoted breakthroughs in architectural efficiency and large - model industrialization.

On the other hand, the part of Alibaba Cloud's ecological strategy that lowers the threshold and shares standards also reflects a sense of technological responsibility. First, the model of "full - stack + open - source + referable industry solutions" has solved many pain points for enterprises in AI implementation, enabling small and medium - sized enterprises to quickly convert the capabilities of the base model into productivity.

Second, an active open - source ecosystem is a devastating blow to black - box operations, monopolies, and unreasonable technological premiums. If we compare AI to the steam engine in the Fourth Industrial Revolution, Alibaba Cloud provides not only the engine and the railway network leading to various application scenarios but also a large number of publicly available design drawings.

In this ecosystem, the "Qwen Architecture" is becoming the de - facto global standard in the AI industry. In 2025, a phenomenon of "reverse learning from China" emerged in the global large - model field: Silicon Valley giants and startup stars, once regarded as technological beacons, began to secretly use Chinese open - source models as a starting point for innovation.

The CEO of Airbnb publicly stated that their core customer - service agents mainly rely on the Qianwen series of models; the flagship model Athene - V2 released by the well - known Silicon Valley AI startup Nexusflow has also publicly stated that it is fine - tuned based on Qwen - 2.5 - 72B. This "export of technology that is more popular abroad than at home" is more intuitively reflected in the developer community Hugging Face: for a long time, most of the top 10 models on the global open - source model ranking