A Chinese on-device AI model becomes the first non-Google AI on Samsung flagships

Samsung's London Unpacked on July 22 will be the first major Android launch in which the company hosts more than one AI engine on its flagship phones. The non-Google half of that stack is now a Chinese model: ModelBest, a Beijing on-device LLM company, has agreed to deploy its self-developed MiniCPM series on several Samsung flagships, per a 36Kr exclusive.

On-device LLM means a model that runs directly on the phone rather than in a remote data center. That is the layer Samsung is now opening to partners, and where a vendor from outside Google's orbit can claim shelf space inside Samsung's global distribution.

The same day 36Kr reported the partnership, China's Cyberspace Administration cleared seven phone on-device generative AI services for public release: Apple Intelligence, Huawei Xiaoyi, OPPO AndesGPT, vivo BlueLM, Xiaomi HyperAI, Samsung Galaxy AI, and Nubia. Samsung's own Galaxy AI is in that list, which means the same group opening up to ModelBest for the global flagship line is also clearing its own on-device engine for Chinese release.

The two events are the commercial and regulatory halves of the same move. Chinese on-device LLMs are crossing the demo-stage threshold inside a single 24-hour window.

In a March 9, 2026 statement reported by Gelonghui, Samsung Electronics president Roh Tae-moon framed the phone maker's strategy as running multiple AI engines on its devices to weaken Apple's lead. That strategy needs partners, and ModelBest is the first publicly named non-Google one for a flagship line. Samsung's multi-AI posture creates room for a co-tenant slot in its distribution stack, and ModelBest has claimed it. What comes next depends on what the model actually does on the device.

ModelBest's pitch is the bet that smartphone-grade silicon gets good enough, fast enough, to make a cloud connection optional for most assistant-style work. The company pivoted fully to on-device in the second half of 2023, when most Chinese peers were still chasing cloud parameter scale. The technical anchor is the 2024 Densing Law paper, written with Tsinghua University, which argues that the intelligence a model can pack into a given parameter budget roughly doubles every 3.5 months, so the parameter count needed for a given level of intelligence drops exponentially. Translated: the same model quality on a phone in 2027 should need a fraction of the parameters it needed in 2024.

The company has shipped a stack of MiniCPM open models that, per 36Kr, have passed 38 million cumulative downloads on GitHub and Hugging Face. The flagship today is MiniCPM-o4.5, a 9-billion-parameter model that handles voice, video, and text in full-duplex, the mode where the model can listen and respond at the same time the way a human does in conversation. It runs on chips from Qualcomm, MediaTek, Intel, Rockchip, NVIDIA, and AMD, plus the Chinese silicon stack, including Huawei Ascend and Cambricon. ModelBest also built BitCPM-CANN, a quantized variant that represents each weight with three possible values rather than the usual sixteen, shrinking model size and memory use. 36Kr reports the trick cuts inference memory roughly sixfold. The 38-million-download and 6x-memory numbers, like the funding figures, are 36Kr's reporting and have not been independently confirmed.

36Kr reports ModelBest raised more than 5 billion yuan (about $700 million) in the first half of 2026, at a valuation above 20 billion yuan (roughly $2.8 billion), the highest among China's on-device AI unicorns. The company was founded in August 2022, incubated from Tsinghua's NLP lab; the co-founders are Li Dahai and CTO Zeng Guoyang, who trained CPM-1, China's first homegrown LLM, as a 22-year-old. Beyond phones, ModelBest's SuperMate on-device agent is expected to ship in more than 300,000 production vehicles by the end of 2026 across Geely, SAIC Volkswagen, GAC, and Mazda, which gives the company another distribution lane to point to when the Samsung deal is announced.

What neither side has answered is the question that will determine whether the partnership is a marquee logo or a structural shift: what does ModelBest actually run on the phone, and where does it sit in a stack that already has Google Gemini? Samsung's July 22 Unpacked invitation bills the event as a "new shape" and lists the venue as London, with global livestream. If the ModelBest integration is more than a press release, the stage demo will show it. If it is a regional fallback that runs only when Gemini is unavailable, the demo will not. That distinction is the story to watch.