Original Title: The Battle of Edge AI Officially Begins
By Xie Zefeng, Edited by Yang Xuran
Stepping away from the fiercely competitive and highly capital-intensive super cloud foundation models, shifting towards edge model R&D, or developing smart terminals with edge models, has become a crucial direction for many enterprises to deeply participate in the AI wave, especially for hardware companies.
Li Dahai, CEO of ModelBest, believes that 2026 will be the "first year of edge AI commercialization." Although the company has not obtained relevant qualifications for AI smartphones, it has also secured substantial financing due to related concepts and products.
In July this year, the Cyberspace Administration of China (CAC) required the orderly filing of AIGC services and added filing information for seven foundation models, including Apple Intelligence, Huawei Xiaoyi, and Xiaomi HyperAI.
Subsequently, the Doubao AI smartphone NaviX Ultra officially debuted on September 16. Apple Intelligence + the new Siri AI has been released at WWDC26, and the recently launched Xiaomi 18 Fold comes with the built-in Xiaomi MiMo edge foundation model.
In the smart vehicle sector, StepFun, led by Yin Qi, is placing bets on both the automotive and smartphone tracks. On the one hand, it has joined hands with Geely to enter the smart vehicle market; on the other hand, it has launched its first native Agentic Phone, the STEPX Neo.
Across the ocean, giants such as Meta, NVIDIA, Google, and OpenAI have also begun to explore edge R&D or collaborate with device manufacturers to create entirely new AI hardware.
In fact, a battle for edge AI has already begun, and the smart hardware industry is facing a massive wave of upgrades.
This article is an in-depth value piece from the content team of "WAVE". Welcome to follow us across multiple platforms.
Achieving More with Less
Compared to widely known cloud foundation models such as DeepSeek and ChatGPT, edge models are more lightweight, have fewer parameters, and can run directly using on-device computing power without relying on the cloud.
The number of parameters is the most significant difference between AI foundation models and edge foundation models. AI foundation models often have trillions of parameters; Kimi K3 (by Moonshot AI) reaches up to 2.8 trillion (2.8T) MoE, and Anthropic's Mythos has up to 10 trillion parameters, with a training data volume of 300 trillion tokens.
The massive number of parameters enables various foundation models in the market to possess stronger performance, capable of capturing complex data patterns and excelling in multiple tasks. However, the problem lies in the fact that relying on cloud-based inference requires a network connection, which brings latency and privacy issues. Once disconnected from the network or encountering unexpected situations, the system will directly crash and become unusable.
Edge devices have limited chip performance and storage space, so the parameters of edge foundation models must inevitably be compressed. With the help of knowledge distillation, pruning, and quantization, the parameter count of edge foundation models can be reduced to the scale of a few billion (B).
Currently, many mainstream model vendors such as DeepSeek, Meta Llama, and IBM Granite have launched small model versions, and their performance and test results for specific tasks are outstanding.
This means that large foundation models can be scaled down to smaller, more lightweight versions to achieve faster inference speeds, lower memory usage, and lower power consumption while maintaining a high level of performance. This means they can be better deployed on terminals such as smartphones, PCs, automobiles, and robots.
Unlike the internet, AI foundation models cannot achieve cost reduction through economies of scale, which is the biggest bottleneck on their path to commercialization.
Training a super foundation model consumes massive resources; the estimated training cost for Mythos alone reaches USD 10 billion. After being put into use, the model's inference cost will climb as the number of daily active users and usage frequency increase.
According to calculations by Guolian Minsheng Securities, including Doubao's free C-end (consumer) services, ByteDance's daily cost is as high as CNY 132 million to CNY 240 million. The extremely high cost of cloud inference will severely delay the sustainability of foundation model commercialization.
Edge AI can "achieve more with less," offering high immediacy and strong privacy. It can also combine local data for more customized innovations, further reducing costs for developers and promoting the commercial implementation of edge AI.
Driven by the "Moore's Law" of chips and the "Density Law" of edge AI, edge intelligence is also accelerating its commercialization.
At the end of 2024, the team of Sun Maosong and Liu Zhiyuan from Tsinghua University, along with the open-source community OpenBMB, proposed the concept of "capability density" to measure the intelligence level per unit parameter of foundation models.
The conclusion of the Density Law is that "the capability density of foundation models grows exponentially over time, approximately doubling every 3.3 to 3.5 months," meaning that "every approximately 100 days, a model with only half the number of parameters can achieve the current optimal performance." Combined with Moore's Law in the semiconductor field, where chip computing power doubles approximately every two years at the same price, the capability density of models doubles approximately every 100 days. Under the resonance of hardware and software routes, AI foundation models are beginning to sink towards terminal devices.
Currently, AI smartphones and AI PCs are accelerating their penetration (such data has certain statistical criteria issues). The penetration rate of generative AI smartphones reached 36% in 2025 and will exceed half by 2027; edge AI will shift from a high-end selling point to a standard product feature. The shipment proportion of AI PCs has climbed to 31% in 2025 and will become the norm by 2029.
Coupled with the form innovation of smart vehicles, embodied intelligence, and even smart hardware, the era of edge AI is imminent.
The Moment of Breakthrough
The emergence of a "duck" is reshaping the global physical AI era and the evolution of edge AI.
On September 3, NVIDIA spent a hefty USD 12.93 billion to acquire Hugging Face, the world's largest AI open-source platform. Just a week prior, Pollen Robotics, under Hugging Face, released a bipedal robot named Microduck priced at USD 399.
This robotic duck, standing only 25 cm tall and weighing less than 800 grams, may have already leveraged the ChatGPT moment for edge AI—orders exceeded USD 2.6 million within 24 hours of opening pre-sales, with an average of one unit sold every 4 seconds during peak periods. If a consumer places an order now, it will take more than 6 months to receive the product.
Microduck adopts an open-source design, allowing users to retrain it themselves, which means this robotic duck is not an ordinary toy but can continuously learn on its own. Players can upload their training results to the Hugging Face Hub, and the massive amount of data generated by countless devices in real environments will drive algorithm evolution at nearly zero marginal cost.
Since the advent of ChatGPT, the narrative thread in the AI world has always been about larger-scale models, more powerful and expensive GPUs, and more massive computing clusters.
Microduck is not equipped with high-end chips from NVIDIA, AMD, or Qualcomm, but rather a mid-range SoC RK3566 developed by Rockchip in 2020.
Compared to humanoid robots that cost hundreds of thousands, such a "duck" with a low entry barrier, fully open-source, and rich ecosystem has significantly lowered the threshold for entering the edge intelligent device market.
As Clem Delangue, CEO of Hugging Face, said, the launch of Microduck signifies that the "era of affordable open-source robots" is arriving, with the goal of enabling more people to participate in the development of physical AI and world models.
On the AI PC front, chip giants such as AMD, Intel, and NVIDIA are accelerating their layouts. In January, AMD launched the Ryzen AI 400 and Ryzen AI PRO 400 processors for Windows 11 AI PCs, with computing power reaching up to 60 TOPS, capable of running foundation models locally.
Intel released the third-generation Core Ultra processor at CES 2026, which is the first computing platform built on the Intel 18A process.
In June, Jensen Huang unveiled the super chip RTX Spark for AI PCs at the GTC conference in Taipei. This chip features 1 Petaflop of computing power and up to 128GB of memory, capable of running a 120-billion-parameter foundation model locally. This autumn, major PC manufacturers such as ASUS, Dell, HP, and Lenovo will successively launch products equipped with this chip.
Driven by multiple factors such as cost, commercial implementation feasibility, and technological evolution, the future evolution of AI foundation models may gradually be divided into three major categories:
The first is the widely known cloud foundation model, which undertakes complex reasoning, long text/video generation, and Agent capabilities, defining the upper limit of AI capabilities;
The second is the industry model, which has expertise in vertical fields such as finance, healthcare, and judiciary, serving professional scenarios and solving AI implementation problems;
The third is the focus of our current discussion, the edge model, whose role is to handle high-frequency, real-time, privacy-sensitive, and offline-available tasks, determining the reach density and breadth of future AI.
The hierarchical synergy of "cloud defines the upper limit, industry defines the depth, and edge defines the entry point" has become increasingly evident.
The Chinese Model
Under the evolution trend of AI foundation models, China's edge AI is also ushering in an explosive singularity.
Among the major players, ModelBest positions itself as a pure edge model foundation supplier, adhering to the concept of "achieving more with less" and attempting to take over all smart hardware with its models; the Doubao smartphone is jointly created by ByteDance and ZTE Nubia to seize the super entry point of edge-cloud collaboration; StepFun defines itself as a native multimodal terminal solution provider, mainly building AI brains for automobiles and smartphones; smart driving solution providers such as Horizon Robotics are software-hardware integrated intelligent foundations, with barriers in automotive-grade chips and software-hardware decoupling capabilities.
Among them, ModelBest, co-founded by Liu Zhiyuan, is an unavoidable enterprise. In the second half of 2023, the company made a "counter-consensus" strategic judgment, abandoning the hundred-billion-level cloud foundation and going all-in on edge models, while originally proposing the "Density Law." Relying on the MiniCPM "Little Steel Cannon" open-source model, it builds an edge model ecosystem.
On September 8, it open-sourced the new generation "Little Steel Cannon" MiniCPM5-2B, with a parameter scale of 2B, focusing on local operation on terminal devices such as smartphones and PCs, while also possessing strong Agent capabilities.
According to the company itself, in the latest evaluation of the AA leaderboard, it achieved the first place among global open-source foundation models with under 4B parameters with a comprehensive score of 23, meaning it achieved the capabilities of a 4B-scale model using only half the number of parameters.
In terms of commercial implementation, ModelBest positions itself as an independent third-party vendor, currently focusing mainly on the smartphone, automotive, and embodied intelligence fields. Among the 7 smartphone edge AI filing lists published by the CAC, the edge capabilities of Samsung's "Galaxy AI" are provided by MiniCPM. Breaking into Samsung's supply chain means that it is not only the first domestic edge AI supplier to enter an international smartphone giant but also the only independent and pure edge model enterprise among the first batch of filed smartphone model companies.
Previously, the automotive smart cockpit assistant cpmGO developed by this company has been mass-produced and landed with Changan Mazda MAZDA EZ-60; in the direction of embodied intelligence, MiniCPM-Robot cuts in with 1.5B parameters and has jointly landed scenarios such as exhibition hall guidance and park inspection with Leju Robotics.
After Yin Qi took charge of StepFun, it accelerated its march towards commercialization. StepFun seems to lean more towards the "software-hardware integration" full chain.
On July 13, StepFun released its first foundation model native AI terminal brand STEPX, and its claimed native Agentic Phone STEPX Neo also made a simultaneous debut. Yin Qi's goal is to connect the complete chain of foundation models, agent systems, and hardware terminals to realize "AI + terminals."
Previously, StepFun, Afari Technology, and Geely have reached a strategic alliance. Afari Technology has also jointly developed the in-vehicle version of Step AOS with StepFun, claimed as the native smart driving foundation for vehicles, deeply integrating edge foundation models, world models, and autonomous driving. In the future, Geely Automobile is highly likely to be the first to implement it.
Among foundation model vendors, Doubao has the most aggressive moves in developing edge AI, extending its tentacles to the smartphone field the earliest.
After testing the waters with the first-generation engineering machine Nubia M153, Doubao and ZTE Nubia adjusted the technical route from the first-generation cross-App solution to the second-generation edge MCP, attempting to create a truly AI Native smartphone.
Nubia NaviX Ultra is claimed to be the "world's first mass-produced AI agent smartphone." It is completely different from simply downloading a Doubao APP. By operating the Agent via voice, it achieves ordering a ride or ordering milk tea with a single sentence. In addition, it also attempts to give AI "eyes" to perceive the physical world.
However, the smartphone still prioritizes coverage of ByteDance's own products, and can also use a small number of third-party applications such as Cao Cao Mobility. Due to security concerns, its Agent functions are not connected to applications such as Alipay and WeChat Pay.
In the competition for C-end AI applications, Doubao has sufficient user accumulation. By jointly developing AI smartphones, ByteDance's open strategy is to continue consolidating Doubao's C-end user base on the smartphone edge and strive for market space in edge AI.
Among smart vehicle solution providers, Horizon Robotics carries out edge deployment around the two major scenarios of smart driving and embodied intelligence. Its advantage lies in software-hardware integration, possessing the Journey smart driving chip + Starry Sky cockpit chip + HSD (Horizon Full-Scenario Driving Assistance System), while also developing the HoloMotion and HoloBrain models for embodied intelligence.
Desay SV, on the other hand, strongly advocates "edge-cloud collaboration." At this stage, excessively high costs still hinder the large-scale installation of foundation models in vehicles. Therefore, when facing diverse customer needs, Desay SV will match different chip computing powers and provide differentiated solutions, including both lightweight edge models and evolutionary plans for the iterative upgrade of full-modal Omni models.
iFLYTEK open-sourced two edge foundation models, Spark X2.5-4B and 1.7B, on September 1. Both can support up to 1 million Token context. Its implementation scenarios focus on "AI office," which not only enables "conversation even when disconnected from the network" but also strengthens the ability to handle complex tasks, with the goal of penetrating into enterprise productivity links.
Final Words
When the competition among foundation models evolves from the "Battle of a Hundred Models" to the "Competition among Top Players," and when top foundation model enterprises seize the dual dividends of capital and the market, the edge has become another focal direction of attention for AI model enterprises.
Low cost, lightweight, offline availability, and enhanced privacy and security are all significant advantages of edge models, presenting the opportunity to usher in a critical moment of commercialization explosion in the near future.
To date, edge AI has not yet demonstrated its unique value and existential significance. As mentioned earlier, this situation will gradually manifest as the functions and roles of foundation models gradually differentiate.
However, an edge model enterprise independent of all third parties and taking all hardware enterprises as hypothetical customers will almost inevitably be disappointed that all potential customers will explore building their own small edge models. From this perspective, the ultimate development of "pure-blooded edge intelligence" enterprises represented by ModelBest will inevitably be limited.
Edge-cloud collaboration may become a trend in the future. Major hardware institutions will explore model compression and the continuous iteration of NPU chips, with the edge gradually supporting Omni full-modal capabilities, and AI sinking into various smart hardware. After such comprehensive penetration of artificial intelligence capabilities, the era of AI inclusive intelligence will be getting closer and closer.