BIP NYC

collapse
Home / Daily News Analysis / Start-Ups Shift AI Tasks To Smartphones

Start-Ups Shift AI Tasks To Smartphones

Oct 24, 2026  Twila Rosenbaum  89 views
Start-Ups Shift AI Tasks To Smartphones

Beijing-based artificial intelligence start-up ModelBest has closed a C+ Series funding round, bringing the total it has raised this year to 5 billion yuan, or about $740 million and £550 million, investors said. The financing highlights growing interest in technology that can run powerful AI tasks directly on devices such as smartphones and laptops instead of sending every request to a remote data center.

Shanghai Mejoy Capital, an investor in the company, said in a social media post that ModelBest completed the round last week. It said the financing valued the on-device AI specialist at 20 billion yuan but did not disclose the size of the latest C+ round. The valuation places ModelBest among a new group of companies betting that the next phase of AI competition will not be confined to giant cloud clusters.

Key facts at a glance

  • ModelBest closed a C+ Series round and has raised a total of 5 billion yuan this year, equivalent to about $740 million or £550 million.
  • Investor Shanghai Mejoy Capital valued the Beijing-based company at 20 billion yuan.
  • ModelBest is a prominent player in on-device AI, a field focused on running models locally on phones, laptops and other edge devices.
  • Shanghai-based StepFun, backed by Tencent Holdings, announced Step Edge on-device models and the StepX Neo smartphone built around autonomous AI agents, without giving a launch date or pricing.
  • Apple has introduced on-device AFM 3 Core and a larger advanced model that activates between 1 billion and 4 billion parameters depending on the task.
  • Alibaba and Baidu are supplying technology for Apple's AI products in China.

Why AI is moving closer to the user

Most advanced AI services today depend on cloud infrastructure. A user types a prompt, the request travels to a data center, a large model processes it and the answer returns over a network. That architecture makes it possible to access very large models, but it also adds latency, consumes bandwidth, creates recurring inference costs and raises privacy questions when personal or business data leaves the device.

Edge AI, also called local AI, takes a different approach. It runs models directly on the handset, laptop or another device, using local processors and memory. The goal is not necessarily to replace the largest cloud models. Instead, companies are trying to move a growing share of everyday tasks, such as summarization, translation, image editing, voice transcription and simple agents, onto local hardware. Those tasks can then be completed faster and more privately, while cloud resources are reserved for heavier work.

The shift is technically demanding. Smartphones and laptops have limited memory, battery life and thermal headroom compared with data centers. Developers must compress models through techniques such as quantization, distillation, pruning and efficient architectures. They also need to coordinate local and remote processing so that the user experience remains seamless. The result is a hybrid model landscape in which small and medium-sized models handle routine requests on-device, while frontier models remain available through the cloud when needed.

China's domestic chip angle

In China, edge AI is attracting attention for another reason: it creates opportunities for locally produced AI chips. Smaller on-device models can be trained and run on hardware that is better suited to lighter workloads than the most advanced frontier systems. That alignment between model design and domestic silicon is becoming an important part of the country's AI strategy as companies seek to build more self-reliant technology stacks.

ModelBest has already pointed to progress in this area. In May, the company said its billion-parameter on-device MiniCPM5-1B model was trained entirely on Huawei's Ascend chips. The model achieved top benchmark scores on Artificial Analysis' index for open-weights models under 2 billion parameters. The result was notable not only for its performance in a compact size class but also for demonstrating that a meaningful on-device model could be developed on domestic hardware.

That matters because the economics of AI are shifting. Training frontier models remains enormously expensive, but inference, the process of running models for users, can account for a large share of ongoing costs. If more inference happens locally, companies can reduce server loads and network expenses while giving users faster responses. For device makers, better local AI can become a selling point that differentiates phones and laptops in a crowded market.

StepFun bets on agents and a new handset

Shanghai-based StepFun, which is backed by Tencent Holdings, is taking the idea further. Last week the company announced Step Edge, a family of on-device AI models designed to offer low latency, local data processing and the ability to handle simple tasks without a network connection. The pitch is straightforward: users should be able to get useful AI assistance even when connectivity is poor, and sensitive data should not have to leave the device for basic operations.

A day later, StepFun announced the StepX Neo, a smartphone designed around autonomous AI agents. According to the company, the agents are built to interpret complex user requests and carry out tasks independently. StepFun did not provide a launch date or pricing for the device. The announcement nonetheless signals a broader ambition among AI companies: rather than treating the phone as a container for apps, they want the phone to become an agent-driven computing platform.

That vision raises both opportunities and questions. An on-device agent could coordinate calendars, messages, travel bookings and other services with less manual input. But it would also need clear permissions, strong security and reliable ways to handle failures. If agents can act independently, users will expect transparency and control. Device makers and model developers will have to build safeguards into the system from the start, not add them later.

Apple's hybrid approach

Apple is pursuing a similar hybrid strategy. The company relies on third-party providers such as Google and OpenAI for some high-end tasks while also developing its own on-device models. Last month it announced the 3 billion parameter AFM 3 Core and a more advanced 20 billion parameter model that activates between 1 billion and 4 billion parameters based on the user's task. Both are designed to run on devices.

The approach reflects a practical compromise. Running a 20 billion parameter model entirely on a phone would be difficult under current memory and power constraints. By activating only a subset of parameters for a given task, Apple can capture more capability while keeping the system within device limits. The strategy also fits Apple's long-standing emphasis on privacy and integration between hardware and software.

Apple is also evaluating technology from Silicon Valley firm PrismML, which claimed last week to have successfully run Alibaba's 27 billion parameter Qwen 3.6 model locally on an iPhone. Such claims, if they become practical at scale, would mark a significant step for on-device AI. They would suggest that phones could eventually run models far larger than today's typical local assistants, though questions about speed, battery use and sustained performance would remain.

In China, Apple's AI products will use technology from Alibaba and Baidu, the companies said last week. The partnerships underscore how regional requirements and local ecosystems are shaping the deployment of AI features. They also show that the move toward on-device intelligence is not happening in isolation. It is intertwined with cloud services, chip supply chains, operating system design and local market rules.

Challenges remain

The excitement around edge AI does not remove the technical obstacles. Local models still struggle to match the breadth and accuracy of the largest cloud systems. Devices vary widely in processing power, memory and software support, creating fragmentation for developers. Battery life and heat are constant constraints, especially during sustained AI workloads. Security is another concern: a model that processes personal data locally can improve privacy, but it must also be protected from malicious apps and unauthorized access.

Even so, the direction of travel is clear. Investors are funding companies that can make models smaller, faster and more efficient. Device makers are adding neural processing units and dedicated AI hardware. Software platforms are building tools to route tasks between local and remote systems. The result is likely to be a gradual blending of cloud and edge AI rather than an abrupt replacement of one by the other.

ModelBest's funding round, StepFun's handset concept and Apple's on-device model family all point to the same competitive question: how much intelligence can be delivered at the point of use without sacrificing capability, privacy or cost. Alibaba and Baidu are providing technology for Apple's AI products in China, the companies said last week.


Source: Silicon UK News


Share:

Your experience on this site will be improved by allowing cookies Cookie Policy