Key Facts

• Beijing-based AI start-up ModelBest has closed a C+ Series funding round, bringing total funds raised 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 smartphones, laptops and other devices rather than sending every request to remote data centres.
• In May, ModelBest said its billion-parameter on-device MiniCPM5-1B model was trained entirely on Huawei Ascend chips and achieved top benchmark scores 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 that offer 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 can interpret complex user requests and carry out tasks independently. No launch date or pricing was given.
• Apple has announced AFM 3 Core, a 3 billion parameter model, and a more advanced 20 billion parameter model that activates between 1 and 4 billion parameters depending on the user task. Both are designed to run on devices.
• Apple is also evaluating technology from Silicon Valley firm PrismML, which said it successfully ran Alibaba's 27 billion-parameter Qwen 3.6 model locally on an iPhone.
• Alibaba and Baidu are providing technology for Apple's AI products in China.

Why On-Device AI Is Attracting Capital

For years, the most capable AI systems have lived in massive cloud data centres. Training a frontier model can require thousands of specialised chips, enormous power supplies and advanced cooling. Running those models for users also means sending data across networks, which introduces latency, adds operating costs and raises privacy questions. On-device AI offers a different path. It moves at least some of the computation onto the phone, laptop or other edge device in the user's hands.

The appeal is straightforward. Local processing can respond in milliseconds because a request does not have to travel to a distant server and back. It can work when a device is offline or when connectivity is poor. It can keep sensitive information such as voice recordings, messages, photos and health data on the device. It can also reduce cloud bills for companies that would otherwise pay for every inference request. For consumers, those benefits can translate into faster features, longer battery life and stronger privacy.

The trade-off is technical. Smartphones have far less memory, compute power and battery capacity than data-centre servers. A model with trillions of parameters cannot simply be copied onto a handset. Instead, companies use smaller models, compression techniques, quantisation, distillation and selective activation. Some systems split work between the device and the cloud, handling simple or private tasks locally while routing heavier requests to remote servers. That hybrid approach is becoming common as AI features move into operating systems, keyboards, cameras, assistants and productivity apps.

ModelBest and the China Edge Ecosystem

ModelBest's funding round is a sign that investors see on-device AI as more than a research curiosity. The company has raised a total of 5 billion yuan this year, according to its investors, and Shanghai Mejoy Capital said the latest round valued it at 20 billion yuan. That valuation places it among the more valuable private AI start-ups in China, even as the broader market for large language models becomes more crowded and expensive to compete in.

ModelBest's strategy reflects a broader trend in China. While companies continue to develop very large models with hundreds of billions or trillions of parameters, they are also building lighter-weight models that can run on local hardware. In China, edge AI is seen as a way to take advantage of domestically produced AI chips. Those chips may not match the most advanced frontier training accelerators, but they can be well suited to inference and to training smaller models. The result is an ecosystem that pairs local silicon with local software.

In May, ModelBest said its billion-parameter on-device MiniCPM5-1B model was trained entirely on Huawei Ascend chips. The company said the model achieved top benchmark scores on an index for open-weights models under 2 billion parameters. That claim matters because it suggests that capable AI models can be developed and run without relying exclusively on the highest-end imported hardware. It also shows how Chinese AI firms are adjusting their engineering priorities around the chips they can access.

On-device models are not limited to phones. Laptops, tablets, cars, cameras, wearables and industrial devices are all potential hosts. A factory camera might use a local model to detect defects without sending video to the cloud. A car might use one to process voice commands or sensor data in real time. A laptop might use one to summarise documents, transcribe meetings or generate code while offline. The common thread is that the model must be small enough, fast enough and efficient enough to run within tight constraints.

StepFun Pushes Agents to the Phone

Shanghai-based StepFun, which is backed by Tencent Holdings, is taking a similar bet. It recently announced Step Edge, a family of on-device AI models designed for low latency and local data processing. According to the company, these models can handle simple tasks without a network connection. That capability could be useful for voice assistants, translation, text summarisation and other everyday features where users expect an immediate response.

A day later, StepFun announced the StepX Neo, a smartphone built around autonomous AI agents. The company described the device as designed for agents that can interpret complex user requests and carry out tasks independently. Instead of merely answering a question, an agent might book a service, organise files, compose messages or coordinate across apps. The company did not provide a launch date or pricing, which suggests the product is still at an early stage.

The idea of an AI agent phone is ambitious. Today's assistants often require precise commands and still fail at multi-step tasks. An agent that runs locally would need to understand context, access device functions, manage privacy permissions and recover from errors. It would also need to do so without draining the battery or overheating the handset. Even if the first generation is limited, the announcement signals where some start-ups believe the market is heading: from AI as a feature to AI as an operating layer.

Apple's Hybrid Approach

Apple is pursuing a similar mix of on-device and cloud AI. The company has announced AFM 3 Core, a 3 billion parameter model, and a more advanced 20 billion parameter model that activates between 1 and 4 billion parameters depending on the user task. Both are designed to run on devices. The selective activation approach is important because it allows a larger model to exist on the device while only using a fraction of its capacity for simpler requests. That can save power and memory while still improving quality for harder tasks.

Apple is also relying on third-party providers such as Google and OpenAI for high-end tasks, while continuing to develop its own on-device models. This hybrid strategy recognises that no single device can run every AI workload. Some requests are best handled locally for speed and privacy. Others require the scale of cloud infrastructure. The challenge is deciding where to draw the line and how to make the transition seamless for users.

Apple is also evaluating technology from Silicon Valley firm PrismML, which said it successfully ran Alibaba's 27 billion-parameter Qwen 3.6 model locally on an iPhone. If such a feat can be repeated reliably in production, it would mark a significant step for edge AI. Running a 27 billion parameter model on a phone requires aggressive compression, memory management and hardware acceleration. It also raises questions about speed, battery impact and how much of the model is actually active at once.

In China, Apple's AI products are expected to use technology from Alibaba and Baidu, the companies said. That arrangement reflects both regulatory requirements and the need for local language and cultural context. It also shows that the on-device AI market is not a single global stack. Different regions may have different model providers, chips and services, even if the underlying trend is similar.

The Road Ahead for Edge AI

The shift toward on-device AI is being driven by several forces at once. Users want faster responses and better privacy. Companies want to reduce cloud costs and differentiate their hardware. Chipmakers want to sell new neural processing units and AI accelerators. Start-ups want to prove that smaller, specialised models can compete with larger cloud systems on everyday tasks.

There are still major obstacles. Memory bandwidth is often the limiting factor on mobile devices. Battery life remains a concern when models run frequently. Developers must optimise for many different chips and operating systems. Model updates and security patches must be delivered without disrupting the user experience. And regulators are paying closer attention to how AI systems handle personal data, which could favour local processing but also create new compliance demands.

Even so, the direction of travel is clear. The first wave of generative AI was defined by cloud-scale models. The next wave may be defined by how well those capabilities can be compressed, distributed and embedded into the devices people carry every day. ModelBest's funding, StepFun's agent phone and Apple's hybrid models are all part of that transition. The competition will not only be about who has the largest model, but also about who can make the smartest use of the smallest amount of compute.

Source: Silicon UK News