HomeArticle

Offline + Memory: The Watershed of Large Model Evolution

晓曦2025-07-27 10:14
The cornerstone of the next-generation general artificial intelligence: The memory mechanism of large models.

This year's WAIC is still as popular as ever.

The sweltering Shanghai, crowded exhibition booths, and clustered AI large models - these have been the norm at WAIC over the years, but there are some changes this year.

There is a notable indicator at this year's WAIC: Whether it's the participating enterprises or the visitors, people have disenchanted the concept of large models. Instead, everyone is asking: "Are there any real - world application cases?"

After three years of market education, users have long been tired of the repeated "model leaderboard rankings". Especially since the beginning of 2025, DeepSeek emerged suddenly, and a large number of AI "star projects" failed one after another. The startup stars are anxious, and so are the investors.

Is there anyone who isn't anxious? There really is.

According to 36Kr, there is a Shanghai - based AI large - model startup that has "quietly" sold its model technology all over the world, with a presence in Africa, the Middle East, Europe, America, Southeast Asia, Russia...

It focuses on "offline intelligence", bringing down the deployment boundary of offline large models to the level of "thousand - yuan smartphones", enabling non - flagship, and even entry - level AIPC, mobile phones, drones, and robots to achieve real - time AI computing offline. In regions with extremely unstable network environments such as Europe, Africa, Russia, and the Middle East, this is almost a "must - have" among "must - haves".

This company is extremely low - key in promotion. It rarely does external PR and doesn't conduct large - scale publicity. There was a joking remark within their team: "We don't go for the 'high - end and fancy'; we aim to 'penetrate Huaqiangbei'."

Ordinary people may think it's self - deprecating, but industry insiders would be shocked - What kind of place is Shenzhen's Huaqiangbei? It's the battlefield with the most concentrated technological and electronic firepower in the country and even the world, a place where the miracle of "molding in one day and mass - producing in three days" happens, the main base for "going global", and also the most fiercely competitive battlefield where countless "underwater unicorns" in the electronic technology industry are born.

This startup is called RockAI (Shanghai Yanxin Digital Intelligence). During WAIC 2025, we finally found an opportunity to have an in - depth conversation with it.

01. The Mysterious "Underwater Unicorn"

RockAI was founded in June 2023 and is a holding subsidiary of A - share listed company Yanshan Technology. Its core product is the first domestic general large - model series based on the Yan architecture without the Attention mechanism, which enables the AI large models, known as "computing power giants", to be deployed offline on low - power devices for real - time computing without an internet connection.

The offline intelligence achieved by RockAI is not simply distilling and pruning the "large model" to get a "small model". Instead, it starts from subverting the underlying architecture, taking a different path from OpenAI, Meta, DeepSeek, and any other large - model companies.

During this WAIC, RockAI also launched the latest generation of the Yan 2.0 Preview large model, further expanding its multimodal capabilities to the video field and introducing a memory unit based on neural networks, enabling the model to have autonomous learning ability.

Low power consumption, high performance, offline intelligence, and non - Transformer are the most prominent labels on RockAI, and also the reasons why customers finally choose RockAI.

02. Offline Intelligence

At the MWC (Mobile World Congress in Barcelona) in March this year, the AIPC products deployed with RockAI's large model were exhibited in the brand's exhibition area, and the on - site feedback was "explosive". Distributors from Europe, America, Africa, Russia and other regions almost immediately expressed their intention to sell. Competitors from the neighboring exhibition areas peeked over curiously and asked mysteriously: "Can it really achieve offline operation at this price?"

Actually, overseas users have a very strong demand for offline AIPC.

For example, one of the most popular features of RockAI's products among overseas users is the intelligent meeting assistant, which can help users with real - time text transcription, translation, meeting recording, and post - meeting summaries during online meetings.

It may sound like the daily routine of "office workers" in China, but many overseas enterprises have extremely strict data security systems, and the data of business meetings are not allowed to be sent to the cloud for processing.

At this time, the offline intelligent meeting assistant can be called a "savior".

In addition, the stability of network facilities is also a daily pain point for overseas users. Devices that are used to the high - speed 5G network in China often have "acclimatization" problems overseas. Not only is the network expensive and unstable, and often drops, but it's really frustrating when it spins for 10 minutes and finally returns a network error message.

Even in China, many brand owners and ODM manufacturers have a strong demand for large - model localization and product differentiation upgrades.

For example, AI capabilities such as face recognition, voice recognition, and OCR photo recognition were once exclusive features of flagship products. However, with a series of innovations in the industrial chain such as algorithms, chips, microphones, and camera modules, these features have now become standard on "thousand - yuan smartphones". Many AI functions have even been extended to offline desk lamps, refrigerators, washing machines, and rice cookers, further expanding the boundaries of device intelligence.

This is the inevitable path for all cutting - edge technologies, and large models are no exception.

Currently, the shipment volume of devices equipped with RockAI's models has reached a certain scale. Its customers include not only the consumer electronics brand owners, ODM manufacturers, mobile phone, robot, and automotive chip companies mentioned above, but also many home appliance and XR glasses manufacturers with more extreme requirements for power consumption and performance, who come to seek cooperation. Many of these most sensitive manufacturers are from Huaqiangbei.

All kinds of "AI Invasion of Huaqiangbei" on social media

03. The Memory of Large Models

Today, when the "hundred - model war" has almost completely subsided, and many star startup projects are in a hurry to transform, go public, and also shout the slogan of "subverting the Transformer architecture", RockAI's large model has been sold all over the world. Its huge and rich technological resources and engineering experience have enabled it to outperform its competitors by a large margin, becoming the most low - key but also the most important "underwater unicorn" in the large - model track.

And RockAI's real ambition is far beyond the intelligence of a computer, a mobile phone, or a home appliance device.

What it wants is what all AI practitioners dream of - Artificial General Intelligence (AGI), and memory is the core of this Yan 2.0 Preview. The robotic dog they demonstrated vividly shows what the "memory" ability of a large model is.

In traditional large models, knowledge often relies on external calls (such as RAG). However, the robotic dog equipped with Yan 2.0 Preview achieves "long - term memory" and personalized understanding in the true sense by deeply integrating memory into the model parameters. This kind of memory is not temporary storage, but an intelligent accumulation that evolves gradually with user interaction, enabling the device to have the ability to "understand you" and become an extension of the user's thinking, rather than just a tool.

More importantly, edge - side deployment makes all this happen on the local terminal, ensuring privacy while also improving response speed and data security. Future devices will no longer be machines that passively execute instructions, but "digital brains" with perception, memory, and learning abilities.

The development of human intelligence fundamentally lies in the accumulation of experience and the evolution of memory. Without memory, there would be no learning, understanding, or the formation of personality. Similarly, if large models want to move from being "powerful tools" to "intelligent agents", they must have human - like memory ability. Only when the model can remember the user's preferences, context, and past interactions can it provide continuously evolving services, achieve truly personalized companionship and decision - making support, and endow machines with human - like thinking continuity and growth potential.

RockAI believes that when each device has memory and autonomous learning ability, they will no longer be isolated nodes, but intelligent agents that can cooperate with each other and share experiences. Such a distributed intelligent network will give rise to "collective intelligence", making the overall system far exceed the sum of its individual parts and forming a social - like learning mechanism. They believe this is a key step towards Artificial General Intelligence.

Conclusion: Large Models at the Watershed

Although it's still too early to talk about AGI at this moment, let's focus on today. Large models have reached an industry watershed. On the one hand, the "hundred - model war" has completely subsided, a large number of star projects are in anxiety, and the industry is undergoing reshuffle, contraction, and layoffs.

On the other hand, there are a large number of real demands from device manufacturers, brand owners, and users for AI large models, and a large number of unmet real - world application scenarios.

Do you still remember that in 2017, Huawei, Apple, and Samsung successively launched their first mobile phone AI chips, officially bringing the booming artificial intelligence war from the cloud to the offline end, thus kicking off the nearly 10 - year national AI boom.

Today in 2025, RockAI, which is "growing against the trend" in the industry's anxiety, may reflect the direction of the trend.