WAIC's Most Groundbreaking Highlight: Offline, Non-Transformer AI Models Are Now in Mass Production — AI Commercialization Is Moving Faster Than Expected

Original Article by Guo Haiwei | AI Technology Review

Published on July 26, 2025, 17:45 | Guangxi

    Rock means stone — the most fundamental, resilient, and ubiquitous element of the world.

    RockAI aspires to be the stone of the intelligent era.      

01 Offline Intelligence at Scale

    According to AI Technology Review, within just 18 months of its launch, devices powered by Yan’s large model capabilities have already entered mass production. As a pioneer of non-Transformer architectures, RockAI has quickly become a key player in edge AI model.

    Visit RockAI’s homepage and you’ll immediately see its bold mission statement:Make Every Device Its Own Intelligence

    RockAI co-founder Jiasi Zou emphasized the two key phrases here — every device and its own.

    Every device reflects RockAI’s pursuit of high compatibility, cost-effectiveness, and wide accessibility.

    Its own speaks to truly autonomous intelligence: self-evolving, continuously growing, and capable of rich interaction and collaboration.This reflects not only a divergence from the mainstream Transformer-based path, but also a deliberate choice to pursue a technically rigorous yet non-hype-driven route.

    But Zou disagrees with calling it a narrow path.

 “The edge device is a massive market,” he says. “This is the kind of intelligence that the vast majority of the world truly needs.”

    In a world dominated by developing economies, intelligence must serve people with dignity.

    One dignified approach is offline AI. Offline is the implicit foundation of edge intelligence — it means user data stays local, won’t uploaded to distant, cold, and unfamiliar servers. It resides on the user’s own device — wrapped in a soft silicone phone case or encased in brushed metal.

    But edge is hard. Offline is harder. And offline-on-edge is not enough.

    A path is still a path — and paths must be walked.Since launching Yan 1.3, RockAI’s journey has been filled with pitfalls — in both technology and real-world application. That’s how mass production of Yan began, and how Yan 2.0 Preview came to be.

    Since Yan1.3 release, many have asked: When is 2.0 coming? Why the delay? What will it look like?

    At WAIC, RockAI is offering a preliminary answer:

    Yan 2.0 Preview.

    True to its modest ethos, RockAI does not differentiate generations solely based on performance metrics — unlike the broader Transformer community. Yan 2.0 Preview is RockAI’s fourth-generation product, following versions 1.0, 1.2, and 1.3.

    Yet it is also the closest embodiment of RockAI’s mission to date. Zou told  AI Technology Review, there are two key advances in Yan 2.0 Preview.

    One is significant improvements in visual perception.

    Yan 2.0 performs sparse frame sampling for video input, reducing temporal redundancy. Selected frames are encoded using a visual encoder to generate visual tokens. A Token Merging (ToME) strategy is applied to reduce semantic redundancy, shortening the token sequence length.

    “It’s no longer just image-level understanding — it can now truly comprehend videos,” says Zou.

    The more groundbreaking advancement lies in its autonomous learning capability enabled by synchronized training and inference.

    RockAI introduces a differentiable memory module that supports dynamic storage, retrieval, and forgetting of information. This architecture allows memory to evolve continuously.Through a novel mechanism, memory strength can be dynamically adjusted — using gated updates to retain long-term dependencies while flexibly integrating new knowledge based on input patterns.Yan 2.0 Preview can now learn like a biological brain — forgetting the trivial and preserving the essential.

    Historically, synchronized training-inference mechanisms were tied to enterprise-grade LLM appliances. But RockAI is bringing this paradigm to personal edge devices — a move some might consider radical.After all, this concept is largely unexplored and lacks mature application scenarios. Even with a functioning product, how it will be used remains an open question.

    Still, RockAI is already exploring use cases with toy manufacturers. For example, discussions are underway with leading global toy companies about AI toys capable of self-learning — where children can talk to, command, and grow with their toy, making it a unique and evolving digital companion.

    “We released the preview version to explore broader upstream and downstream opportunities,” says Zou.“We firmly believe in its vast potential.” And vast potential means far more than just toys.

    If Yan 1.0 was a text model, and Yan 1.3 a multimodal model with image and audio capabilities, then Yan 2.0 marks a leap forward — enabling video understanding, autonomous learning, and real-time human-machine interaction. It is a new foundational model for the age of truly personal AI.

02  It’s All About Real-World Scenarios

    Technical pitfallswere just one of RockAI’s two major takeaways from the past year — the other, even more critical, was the challenge ofreal-world scenarios.

    “Scenarios are hard,” said RockAI co-founder Jiasi Zou in an interview with.

    The gap between demos and deployment is significant — and scaling to mass production is even harder.

    For example, deploying Yan1.2 on a Raspberry Pi was once considered a milestone worthy of being written into the company handbook. But following the launch of Yan 1.3, RockAI secured contracts with leading overseas hardware manufacturers — a major step forward for business growth.

    Zou still vividly recalls when a partner showcased a PC powered by Yan 1.3 to overseas distributors — their shock and excitement were palpable. Even representatives from major competitors stopped by to ask how RockAI had achieved offline, on-device LLM inference.

    Offline matters more than people think.

    “There’s enormous demand for offline, edge-deployed large models in global markets,” said Zou. “The challenge is that many of those needs remain unsolved.”

    For hardware companies going global, navigating varied and evolving data privacy regulations is complex. On-device AI avoids costly compliance with cross-border data transfers and cloud-related policies.

    “Privacy is a very sensitive issue overseas,” said Zou. “Different countries have vastly different laws regarding personal data. Uploading user data would mean clearing multiple legal and regulatory hurdles in each region.”

    This is further complicated by inconsistent global internet infrastructure.

    Currently, the majority of the global population lives in developing countries, with rural populations significantly outnumbering urban ones. According to a 2022 report by the ITU, in Europe and North America, 80% to 90% of the population has internet access. However, this figure drops to around two-thirds in Arab and Asia-Pacific countries. In Africa, the percentage of internet users further declines to 40%, while in the least developed countries and landlocked developing nations, internet coverage is only 36%. The urban-rural divide is similarly stark. According to the same report, even in resource-rich cities, only 82% of urban residents have access to the internet at home, which is 1.8 times the rate in rural areas.

    RockAI’s vision is to empower every device, not just every device in Europe, every Chinese device, or every city device.

    It means every device — without qualifiers.

    “Many devices lack NPUs — some don’t even have GPUs. All they have is a CPU. How do you run a large model on that? Only RockAI can do it now days” said Zou.

    In fact, for many global users, Yan might become their very first experience with an AI model.

    That’s why overseas distributors were so surprised by Yan 2.0 Preview running on AIPC devices — they had never seen anything like it.

    “An AI product that runs entirely offline on mid- to low-end CPUs? They hadn’t seen that before,” said Zou.

    To bring this product to market, RockAI invested heavily — not just funding, but technological commitment. This stems from a distinctive go-to-market philosophy: the flagship strategy.

    Unlike companies that chase benchmark scores, RockAI doesn’t follow leaderboard culture.

    “We run benchmarks, but we don’t play to hit the chart,” said Zou.

    Though the team plans to participate in more credible evaluations in the future, the company’s priority remains scenario-first — delivering real-world applications.

    “Back then, Texas Instruments pitched an unknown ARM core to Nokia for small devices. After long negotiations, Nokia said yes — and that partnership helped create both Nokia’s mobile dominance and ARM’s global success.”

    RockAI doesn’t expect a Nokia-style savior, but it does aim to land clear, visible, and deliverable reference scenarios.

    These flagship scenarios are taken seriously. Zou recalls winning a top-tier hardware contract largely because RockAI was the only team that brought a working demo.

    “We threw together a small Core i5 box two weeks before the meeting. The mic and monitor were plugged in on-site. It was messy,” said Zou.

    “But while everyone else had PowerPoints, we had a real device running real AI.”

    The client didn’t even wait for the introduction before expressing interest.“They asked: ‘When was your company founded?’ only after deciding to work with us.”

    Despite being the last vendor to engage with the client, RockAI was the one that won the deal.

    Benchmarks can be manipulated, real deployments can't. User feedback and commercial orders are honest. Device mass production is not just a business milestone — it marks the beginning of the Yan ecosystem.

03 Remain On the Table

    No one doubts that RockAI is battling giants. It’s like David facing Goliath, armed with just a few stones. David used a slingshot to turn those stones into a small yet powerful flywheel.

    The relationship between YAN and the Transformer community isn’t one of absolute competition. It’s more like a race between a river and the ocean’s tides.

    However, for a closed-source large model ecosystem, timing is crucial. This is because RockAI has to balance both foundational model development and application creation; they must manage the upper-layer interactions and lower-layer adaptations while constantly navigating through technical and scenario pitfalls.

    Jiasi Zou mentioned that over the past year, RockAI’s team has grown to nearly 100 people, with the majority focused on R&D.

    For a commercial company, the concept of timing essentially boils down to identifying what to prioritize and what to let go of.

    Zou is reluctant to label RockAI’s large model as closed-source. Instead, he believes RockAI will follow its own open-source rhythm — some aspects may take time to disclose, but they have already made some core components open to key partners, such as chip manufacturers. This approach helps lower the deployment friction for Yan.

    Some initiatives, such as image generation, are temporarily paused. Zou explains that the core requirements for edge devices focus on perception, decision-making, and action.

    “Just like humans,” Zou said, “The human brain doesn’t directly generate images. It uses tools to process information — Yan follows a similar approach.”

    For more visionary projects, which are not yet implemented but are key industry trends, RockAI’s objective is remian on the table.

    Zou sees robotics as a future scenario, but Yan must stay competitive in that race. He believes Yan is naturally compatible with robotics, especially since it focuses on multimodal capabilities centered around perception and cognition rather than generation. In their experiments, RockAI has already demonstrated that Yan’s architecture can replicate basic motion trajectories.

    “At our booth, we showcased a robotic hand that can play games,” Zou said. “We also had a robotic dog that can learn human movements.”

     In RockAI’s vision, it will span a broad spectrum, from UAVs at the high-end to robots and devices like smartphones, computers, and smart glasses, down to companion toys and low-power devices. Each of these devices will be part of Yan’s computational network, contributing to the collective intelligence that will define the future world.

    Who says small stones can’t form great mountains?