Beijing-based artificial intelligence start-up ModelBest has completed a C+ Series funding round, bringing its total fundraising this year to 5 billion yuan, or roughly $740 million and £550 million, according to investor Shanghai Mejoy Capital. In a social media post, the investor said the round valued ModelBest at 20 billion yuan, although the size of the latest round was not disclosed. The financing underlines the growing appetite for technology that can run demanding AI workloads directly on smartphones, laptops and other local devices instead of routing every request through remote cloud facilities.

ModelBest is one of the more prominent companies focused on on-device AI, a field that sits at the intersection of model design, chip efficiency and mobile software. Its funding comes as AI developers pursue two tracks at once: building ever-larger foundation models with trillions of parameters, and creating smaller, more efficient models that can operate locally. The second track, often described as edge AI or local AI, is becoming more commercially important as device makers look for ways to deliver faster, cheaper and more private AI features.

Key facts at a glance

• ModelBest closed a C+ round and has raised 5 billion yuan this year, equivalent to about $740 million or £550 million.
• Shanghai Mejoy Capital valued ModelBest at 20 billion yuan and did not disclose the size of the latest round.
• ModelBest is a major player in on-device AI in China.
• The company's billion-parameter MiniCPM5-1B on-device model was trained entirely on Huawei Ascend chips, according to ModelBest.
• ModelBest said the model achieved top benchmark scores on Artificial Analysis' index for open-weights models under 2 billion parameters.
• Tencent-backed StepFun announced Step Edge, a family of on-device AI models offering low latency, local data processing and simple tasks without a network connection.
• StepFun also announced the StepX Neo, a smartphone built around autonomous AI agents, but did not provide a launch date or pricing.
• Apple announced a 3 billion parameter AFM 3 Core model and a more advanced 20 billion parameter model that activates between 1 billion and 4 billion parameters depending on the task.
• Apple is 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.

The current division of labor in AI is straightforward in principle but difficult in practice. Some AI features already run locally on phone handsets, handling tasks such as image enhancement, speech recognition, text prediction and basic personalization. More compute-intensive chores are typically routed through remote facilities, which adds latency, raises operating costs and creates privacy questions because user data must leave the device. Edge AI seeks to move more of that work back onto the handset, laptop or other endpoint.

For consumers, the benefits can be immediate. A local model can respond without a network round trip, making features feel faster and more reliable in areas with poor connectivity. It can also keep sensitive information, such as messages, photos or health data, on the device. For developers and device makers, local processing can reduce cloud inference bills and help manage network bandwidth. Those advantages are especially relevant for smartphones, which have limited battery life and thermal headroom but are carried everywhere and generate large amounts of personal context.

Why the cloud is not disappearing

None of this means cloud AI is being replaced. The most demanding tasks still require large data-center models, and a hybrid approach is likely to dominate. A phone might use a small local model for quick drafting, summarization or voice commands, then call on a remote model for complex reasoning, long-context analysis or multimodal generation. That routing decision has to balance latency, cost, privacy and quality. Model makers are increasingly designing systems that can decide where a task should run, rather than treating local and cloud AI as mutually exclusive.

The engineering challenge is considerable. Running powerful models on a phone requires compressing and adapting them for tight memory budgets, limited battery capacity and heat constraints. Techniques such as quantization, pruning, distillation and efficient attention mechanisms can reduce model size and compute demands, but they may also affect accuracy. Device makers must coordinate model design with specialized neural processing units, memory bandwidth and operating-system scheduling. The result is a close partnership between AI researchers, chip designers and handset manufacturers.

China's edge AI strategy

In China, edge AI is also seen as a way to take advantage of locally produced AI chips. These chips may be better suited to training and running lighter-weight models than the most advanced frontier products, especially as access to some high-end imported hardware remains constrained. ModelBest's claim that its billion-parameter MiniCPM5-1B model was trained entirely on Huawei's Ascend chips is notable because it shows how domestic chip stacks are being used for smaller, device-oriented models. The company said the model achieved top benchmark scores on Artificial Analysis' index for open-weights models under 2 billion parameters.

That benchmark performance does not automatically translate into a better user experience. Real-world performance depends on how well a model handles specific tasks, how quickly it responds on particular hardware, how much memory it consumes and how gracefully it degrades when a task exceeds local capacity. Still, the result signals that the gap between small local models and larger cloud models is narrowing for a range of everyday functions. As more models are optimized for on-device use, smartphone makers gain more flexibility to add AI features without forcing every interaction through a remote server.

StepFun's AI-agent phone

Shanghai-based StepFun, which is backed by Tencent Holdings, has moved in a similar direction. It recently 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. A day later, the company announced the StepX Neo, a smartphone designed around autonomous AI agents built to interpret complex user requests and carry out tasks independently. StepFun did not give a launch date or pricing for the device.

The concept of an AI-agent phone goes beyond a voice assistant that answers questions. An agent is meant to understand a goal, break it into steps and act across apps or services with less direct supervision. That could include organizing travel, managing messages, scheduling meetings or completing multi-step workflows. Running such agents locally could improve responsiveness and privacy, but it also raises questions about reliability, permissions and user control. A local agent that makes mistakes quickly is not necessarily better than a cloud agent that makes them slowly.

Apple's hybrid approach

Apple is taking a similarly hybrid route. The company relies on third-party providers such as Google and OpenAI for some high-end tasks while also working on 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 user tasks. Both are designed to run on devices. The approach reflects a broader industry pattern: keep common, latency-sensitive and privacy-sensitive tasks local, while sending the most complex requests to larger models in the cloud.

Apple is also evaluating technology from Silicon Valley firm PrismML. The company claimed last week that it had 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 be a significant technical feat, though details such as speed, memory use and sustained performance matter as much as the initial demonstration. Apple's interest suggests that it is looking beyond its own models to third-party compression and deployment techniques that could make larger local models practical.

In China, Alibaba and Baidu are providing technology for Apple's AI products, the companies said last week. That arrangement reflects both the importance of the Chinese market and the regulatory and operational need for local partners. It also shows how the edge AI supply chain is becoming more layered: chip makers, model developers, device manufacturers and platform providers all have roles to play. The competitive landscape is no longer just about who has the largest model. It is also about who can deliver useful AI features within the power, memory and thermal limits of a pocket-sized device.

Cost, privacy and the limits of local AI

The economic logic of edge AI is compelling but not unlimited. Local inference can reduce cloud costs, but it also shifts costs to the device, where battery life and silicon area are scarce. A phone that runs a large model locally may drain its battery faster or heat up during prolonged use. Device makers must therefore decide which features are worth running on-device and which should be offloaded. For users, the best experience may be one in which the transition between local and cloud processing is invisible, with the system choosing the most appropriate location for each task.

Privacy is another central driver. Keeping data on a handset can reduce exposure during transmission and storage, but local processing is not automatically private. Apps may still collect and share data, and on-device models can be used to infer sensitive information. Regulation and platform rules will shape how much data remains local and how clearly users are informed. The technical architecture matters, but so do the policies around it.

The funding and product announcements point to a busy period for on-device AI. ModelBest's valuation and StepFun's device concept show that investors and entrepreneurs see a market beyond cloud-only AI. Apple's model roadmap and its evaluation of third-party deployment technology show that large platform companies are preparing for a more distributed AI stack. The next round of competition will likely focus on which company can make local AI feel genuinely useful, rather than merely demonstrating that a model can run on a phone.

For now, several details remain undisclosed. ModelBest's latest round size was not revealed, and StepFun has not announced a launch date or pricing for the StepX Neo. Apple has not provided a public timeline for the full rollout of its next-generation on-device models or for any potential use of PrismML technology. Those gaps matter because the commercial impact of edge AI will depend on shipping products, real-world performance and pricing, not only on benchmark scores and funding announcements.

Source: Silicon UK News