Key Facts

• ModelBest, a Beijing-based on-device AI start-up, has completed a C+ Series round.
• The company has raised 5 billion yuan so far this year, equivalent to about $740 million or £550 million.
• Investor Shanghai Mejoy Capital said the round valued ModelBest at 20 billion yuan but did not disclose the latest round size.
• ModelBest is known for local or edge AI, which runs AI tasks on devices rather than solely in remote data centers.
• The company said in May that its MiniCPM5-1B on-device model was trained entirely on Huawei Ascend chips.
• The model achieved top benchmark scores on 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.
• StepFun also announced the StepX Neo smartphone, built around autonomous AI agents, without giving a launch date or pricing.
• Apple has announced on-device models including the 3 billion parameter AFM 3 Core and a 20 billion parameter model that activates 1 to 4 billion parameters based on tasks.
• Apple is evaluating technology from PrismML, which claimed to run Alibaba’s 27 billion parameter Qwen 3.6 model locally on an iPhone.

Beijing-based artificial intelligence start-up ModelBest has closed a C+ Series funding round, bringing the total it has raised so far this year to 5 billion yuan, or about $740 million and £550 million, according to investors. The deal underscores a widening shift in the AI industry: companies are moving more powerful tasks from remote data centers onto smartphones, laptops and other edge devices.

ModelBest, one of the more prominent players in on-device AI, completed the round last week. Investor Shanghai Mejoy Capital disclosed the completion in a social media post. Shanghai Mejoy said the round valued ModelBest at 20 billion yuan, though it did not disclose the size of the latest financing. The valuation and the accumulated fundraising total signal that investors see local AI processing as more than a niche feature.

Why AI Is Moving to the Edge

Currently, some AI features already run locally on phone handsets, but more compute-intensive tasks rely on routing chores through remote facilities. That approach adds latency and costs and raises privacy questions. Every request sent to a cloud data center must travel across a network, wait for processing and return a result. For tasks such as photo editing, voice transcription or instant translation, even small delays can make an experience feel slow. Cloud processing also consumes bandwidth and can expose sensitive data to additional security risks.

The emerging alternative is local or edge AI. Instead of sending every task to a remote server, devices run models directly on their own chips. This can reduce latency, keep personal data on the device and allow some features to work without a network connection. The trade-off is that phones and laptops have far less memory, power and cooling capacity than data centers. That has forced AI developers to design smaller, more efficient models and to use techniques such as compression, quantization and selective activation.

The industry is not abandoning large models. Companies continue to develop models with trillions of parameters for the most demanding workloads. At the same time, they are building models and techniques for running increasingly powerful tasks locally. Many systems are likely to use a hybrid approach, handling simple or time-sensitive requests on the device while sending harder problems to the cloud. The balance between local and remote processing will depend on the task, the hardware and the user’s privacy preferences.

ModelBest and the China Chip Angle

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. That makes on-device AI strategically important for a market that is building its own semiconductor supply chain and looking for practical uses for domestic hardware.

In May, ModelBest said its 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 matters because it shows that a compact model can be trained on domestic hardware and still compete on standardized evaluations. It also illustrates the broader push toward smaller models that can fit within the memory and power limits of phones and laptops.

The funding round suggests that investors are willing to back that approach. A 20 billion yuan valuation places ModelBest among the more closely watched on-device AI companies in China. The company has now raised 5 billion yuan this year, a significant sum for a start-up working on local AI. The capital could help it expand research, hire talent and prepare for a market where device makers are looking for AI features that do not depend entirely on cloud services.

StepFun Pushes On-Device Models and Agentic Phones

Shanghai-based StepFun, backed by Tencent Holdings, is pursuing a similar 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. Those features are aimed at everyday interactions where speed and privacy matter, such as voice commands, summarization and routine assistance.

A day later, StepFun announced the StepX Neo, a smartphone designed around autonomous AI agents built to interpret complex user requests and carry out tasks independently. The company did not give a launch date or pricing. The concept goes beyond a conventional voice assistant. An agentic phone would be expected to understand a goal, break it into steps and complete actions across apps and services. Running such agents locally could make them more responsive and keep more personal context on the handset.

The announcement also reflects a broader competition among device makers and AI companies. Smartphones are becoming a key battleground for AI features, but the hardware imposes strict limits. Battery life, thermal performance and memory bandwidth all shape what can run on a phone. Companies that can optimize models for those constraints may gain an advantage in delivering useful AI without draining the battery or relying on constant connectivity.

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 every AI request can be handled locally. Complex reasoning, large-scale content generation and other demanding workloads may still require cloud infrastructure. But Apple has been building a local layer for tasks where speed, privacy and offline access are important.

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 selective activation approach is notable: instead of using the full model for every request, the system can engage only the parameters needed for a given task. That can reduce compute demands and make advanced AI more practical on consumer hardware.

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. If such claims hold up in commercial products, they would mark another step toward putting larger models directly into users’ hands. Running a 27 billion parameter model on a phone would require aggressive optimization, memory management and likely selective activation, but the direction of travel is clear.

The China market adds another layer. Alibaba and Baidu are providing technology for Apple’s AI products in China, the companies said last week. That arrangement highlights how AI features are being localized for different markets, with domestic partners supplying models and services that meet local requirements. It also shows that the shift to edge AI is not happening in isolation. It is intertwined with cloud partnerships, chip supply, model design and regional regulation.

Source:Silicon UK News