ModelBest, a Chinese developer of on-device artificial intelligence models, has reached a partnership with Samsung. Its MiniCPM models will be installed on several of Samsung’s flagship smartphones.
On the day the partnership was announced, China’s cyberspace authorities said seven generative AI services designed for smartphones had completed regulatory registration. The list included Apple Intelligence, Huawei’s Xiaoyi, Oppo’s AndesGPT, Vivo’s BlueLM, Xiaomi HyperAI, Samsung’s Galaxy AI, and ByteDance’s Doubao for Nubia smartphones.
The registrations suggest that on-device AI is moving beyond demonstrations and into wider commercial use.
Alibaba also confirmed that Qwen would be integrated into Apple Intelligence on devices sold in mainland China. The integration will cover iOS, iPadOS, macOS, and visionOS, allowing users to access Qwen’s text and image understanding and content-generation functions without leaving Apple’s apps.
Alibaba shares rose after the announcement.
Chinese smartphone makers have so far relied largely on models developed in-house. Huawei has Xiaoyi, Oppo has AndesGPT, Vivo has BlueLM, and Xiaomi has HyperAI.
Qwen’s integration with Apple Intelligence and ModelBest’s partnership with Samsung suggest that smartphone makers are becoming more willing to work with outside model developers rather than build every AI function themselves. A division of labor is beginning to emerge.
ModelBest was founded in Beijing in August 2022 and incubated by Tsinghua University’s Natural Language Processing Lab.
Co-founder and CEO Dahai Li previously served as a partner and CTO at Zhihu. Co-founder and chief scientist Zhiyuan Liu is a professor in Tsinghua University’s Department of Computer Science and Technology. Co-founder and CTO Guoyang Zeng was twenty-two when, as an engineering lead, he helped train CPM-1, which has been described as China’s first large language model.
According to PEdaily, ModelBest raised more than RMB 5 billion (USD 737.7 million) in the first half of 2026. The publication said the company’s valuation exceeded RMB 20 billion (USD 3 billion), making it China’s most highly valued developer focused on on-device AI.
ModelBest was also among the first Chinese companies to concentrate on the field.
In the second half of 2023, while much of the industry was occupied with increasing the size of cloud-based models, ModelBest turned its attention to models that could run directly on consumer devices.
At the time, the choice appeared unfashionable. Large model developers were competing over parameter counts and computing power, while on-device models were often treated as a secondary field concerned mainly with compression.
ModelBest based its approach on a method it calls “knowledge density.”
In 2024, ModelBest and a team from Tsinghua University proposed what they called the “densing law.” It holds that the maximum capability density of open-source large models doubles about every three and a half months. If the relationship holds, the number of parameters needed to reach the same level of performance should decline exponentially.
Later model releases have offered some support for that assessment.
In May, ModelBest, Tsinghua University, and OpenBMB released MiniCPM5-1B as an open-source on-device text model. Although it has one billion parameters, the model scored 17.9 on the Artificial Analysis Intelligence Index, outperforming several larger open-source models.
MiniCPM-V 4.6 has 1.3 billion parameters and can run on a smartphone with six gigabytes of memory, according to ModelBest. It supports iOS, Android, and HarmonyOS.
The company said MiniCPM’s open-source models have been downloaded more than 38 million times across GitHub, Hugging Face, and other platforms.
In an interview with 36Kr, Zeng described the constraints of on-device development.
“On-device deployment comes with hard constraints. If a model is too large to run, then it simply cannot run, and subsidies will not solve the problem. If power consumption is too high, the device will overheat,” he said.
In Zeng’s view, making a model smaller does not make the work easier.
“Making a model efficient is extremely difficult. Bigger does not necessarily mean simpler,” he said.
Deploying a model on a device involves more than reducing its size. Developers must coordinate algorithms, chip optimization, power use, memory demands, and installation on the final product.
ModelBest has adapted its technology for chip platforms made by Qualcomm, MediaTek, Intel, Rockchip, Nvidia, and AMD.
For Chinese chipmakers, it has helped develop software systems for platforms including Huawei Ascend and Cambricon. The company also released BitCPM-CANN, which it describes as China’s first ternary-quantized large model. ModelBest said the model can reduce GPU memory use during inference by about sixfold.
The company has also developed models for several types of on-device use.
Its latest release, MiniCPM-o 4.5, is an omnimodal model with nine billion parameters. It can process voice, video, and text at the same time and supports full-duplex interaction, allowing a user and the system to speak without waiting for one another to finish.
Running such a model on a smartphone requires careful control of power consumption, heat, memory use, and response time. A gain in one area can create costs elsewhere.
“On-device models cannot be judged solely by benchmark performance. Response time and hardware costs also matter, which makes it difficult for a model to prevail based on any single dimension,” Zeng said in the interview.
ModelBest said it has already tested its ability to deploy its technology at scale in automobiles.
Its on-device AI agent, SuperMate, is expected to appear in more than three hundred thousand production vehicles by the end of 2026, according to the company. The vehicles will include models from Geely, SAIC Volkswagen, GAC Group, Mazda, and other manufacturers.
Through these projects, ModelBest said it has developed a process covering model adaptation, hardware compatibility, supply chains, and delivery to finished products.
Samsung is one of the world’s largest Android smartphone makers by shipments. Its on-device AI functions have so far relied mainly on Galaxy AI and its partnership with Google.
South Korean media have reported that Samsung is working with Google to integrate Gemini across its devices. The reports said the collaboration would include products in the Galaxy Unpacked 2026 lineup, scheduled to be unveiled in London on July 22.
ModelBest’s precise role within Samsung’s AI system has not yet been disclosed.
KrASIA features translated and adapted content that was originally published by 36Kr. This article was written by Deng Yongyi for 36Kr.
Note: RMB figures are converted to USD at rates of RMB 6.78 = USD 1 based on estimates as of July 21, 2026, unless otherwise stated. USD conversions are presented for ease of reference and may not fully match prevailing exchange rates.