Artificial intelligence start-ups are pushing a growing share of AI workloads out of remote data centers and onto smartphones, laptops and other personal devices. The shift is drawing fresh capital and shaping product strategies, as investors bet that local processing can make AI faster, cheaper and more private.

Key facts at a glance

• Beijing-based ModelBest closed a C+ Series funding round, bringing its total raised so far this year to 5 billion yuan, or about $740 million and £550 million.
• Investor Shanghai Mejoy Capital said the round valued ModelBest at 20 billion yuan. The size of the latest round was not disclosed.
• ModelBest is a prominent player in on-device AI, a field focused on running AI tasks locally on devices rather than relying entirely on cloud infrastructure.
• In May, ModelBest said its MiniCPM5-1B on-device model, which has about one billion parameters, was trained entirely on Huawei's Ascend chips.
• The model achieved top benchmark scores on the Artificial Analysis index for open-weights models under 2 billion parameters.
• Shanghai-based StepFun, backed by Tencent Holdings, announced Step Edge, a family of on-device AI models offering low latency, local data processing and the ability to handle simple tasks without a network connection.
• StepFun also announced the StepX Neo, a smartphone designed around autonomous AI agents that interpret complex user requests and carry out tasks independently. No launch date or pricing was given.
• Apple is working on on-device models while relying on third-party providers such as Google and OpenAI for high-end tasks. Alibaba and Baidu are providing technology for Apple's AI products in China.

Funding signals investor appetite

The C+ Series round at ModelBest shows that investors see a commercial opening in AI that runs closer to the user. The company has now raised 5 billion yuan this year, according to Shanghai Mejoy Capital, which announced the deal in a social media post. Shanghai Mejoy also said the round valued ModelBest at 20 billion yuan, although the investor did not disclose the size of the latest financing.

That valuation reflects more than enthusiasm for another AI model developer. It points to a broader bet that the next phase of AI adoption will not be defined only by giant cloud models with trillions of parameters. Instead, many everyday tasks may be handled by smaller, specialized models that live on a phone, laptop or other edge device. Such models can respond quickly, work with sensitive data locally and reduce the constant need to send information to remote servers.

For start-ups, the appeal is strategic. Cloud-based AI can be expensive to operate because every request consumes remote compute, bandwidth and energy. Edge AI can spread some of that work to hardware that users already own. For consumers, the benefit can be lower latency, better offline functionality and stronger privacy protections. For device makers, local AI can become a selling point that differentiates new phones and computers.

Why edge AI is moving to the foreground

Some AI features already run locally on phone handsets. Voice recognition, image processing and simple predictive features often rely on on-device machine learning. But more compute-intensive tasks, including advanced language generation and complex reasoning, typically route chores through remote facilities. That architecture adds latency and costs, and it raises privacy questions because user data may leave the device.

Companies are now trying to move more of that workload back onto devices. They continue to develop large models with trillions of parameters, but they are also developing smaller models and techniques for running increasingly powerful tasks locally. This approach is known as local AI or edge AI. It depends on efficient model architectures, compression, quantization and chips designed to handle AI inference within tight power and memory limits.

The engineering challenge is significant. Smartphones have limited battery life, memory bandwidth and thermal headroom compared with data centers. A local model must deliver useful results without draining a battery or overheating a handset. That is why many companies are designing hybrid systems: simple or privacy-sensitive tasks run on the device, while the most demanding requests are sent to the cloud. The balance between local and remote processing is becoming a central design decision for AI products.

China's chip strategy shapes the edge

In China, edge AI is seen as a way of taking advantage of locally produced AI chips that are better suited to training lighter-weight models than frontier products. The goal is not only to improve consumer devices but also to build a technology stack that relies less on advanced foreign hardware. Domestic chips can support a wide range of AI workloads, especially when models are optimized for efficiency rather than raw scale.

ModelBest's MiniCPM5-1B offers an example. The company said in May that the on-device model was trained entirely on Huawei's Ascend chips. It also achieved top benchmark scores on the Artificial Analysis index for open-weights models under 2 billion parameters. The result is notable because it suggests that smaller models can be competitive within their weight class while being trained on domestic hardware. For China's AI ecosystem, that combination of lightweight models and local chips could make edge AI more practical and more scalable.

The strategy also fits the economics of consumer devices. Not every task needs a frontier model. Many useful features, such as summarization, translation, photo editing, personal assistants and routine automation, can be handled by compact models that are optimized for specific jobs. If those models can run locally on widely available hardware, AI features can reach more users without imposing the same cloud costs on every interaction.

StepFun's on-device push

Shanghai-based StepFun, backed by Tencent Holdings, is moving in the same direction. Last week, the company announced Step Edge, a family of on-device AI models offering low latency, local data processing and the ability to handle simple tasks without a network connection. The pitch is straightforward: users get faster responses and more control over their data, while devices can keep working in places where connectivity is limited or unreliable.

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. The company did not give a launch date or pricing. Even so, the announcement signals a growing interest in hardware that is designed from the ground up for AI agents rather than treating AI as an add-on to a conventional smartphone.

The concept raises important questions for the industry. If agents can act on a user's behalf, they need access to personal context, apps and services. That makes privacy, security and reliability central concerns. Local processing can help by keeping some sensitive data on the device, but it also requires careful system design so that agents do not overreach or consume too much power. The smartphone form factor is a natural testing ground because it combines powerful chips, sensors, cameras and always-on connectivity in a single device.

Apple's hybrid approach

Apple is similarly relying on third-party providers such as Google and OpenAI for high-end tasks, while also working on on-device models. That hybrid strategy reflects the reality that not all AI requests can or should run locally. Some tasks require the scale of cloud infrastructure, while others can be handled efficiently on a phone or laptop.

Last month, the company announced the 3 billion parameter AFM 3 Core and a more advanced 20 billion parameter model that activates between 1 and 4 billion parameters based on user tasks. Both are designed to run on devices. The approach is notable because it uses a larger model selectively, activating only the portion needed for a given request. That kind of conditional computation can help balance capability against efficiency.

Apple is also evaluating technology from Silicon Valley firm PrismML. PrismML claimed last week to have successfully run Alibaba's 27 billion-parameter Qwen 3.6 model locally on an iPhone. Running a model of that size on a handset would represent a significant technical milestone, though real-world performance depends on factors such as speed, memory use, battery impact and the quality of the user experience.

The competitive landscape is becoming more complex. Device makers want to control the AI experience, cloud providers want to supply frontier models, and start-ups want to carve out roles in model design, optimization and chips. In China, Alibaba and Baidu are providing technology for Apple's AI products, the companies said last week. That arrangement illustrates how global device platforms and local AI providers may need to work together as AI features become more deeply embedded in everyday hardware.

Source:Silicon UK News