From a Tsinghua Lab to a 20-Billion-Yuan Unicorn: Mianbi Intelligence’s “Edge Computing” Breakthrough

In the summer of 2026, the edge AI sector welcomed a unicorn company valued at over 20 billion yuan. On July 15, Beijing Mianbi Intelligent Technology Co., Ltd. completed a new Series C+ funding roundfinancingIts investors include a diverse range of industrial players and financial investors, such as national-level funds, state-owned enterprises, and automakers. In the first half of 2026 alone, the company raised more than 5 billion yuan, catapulting it to the position of the startup with the highest publicly disclosed valuation in China’s edge computing sector. Starting from a single office on Zhongguancun East Road in Haidian District in 2022, Mianbi Intelligence has now achieved large-scale deployment across multiple sectors—including mobile phones, automobiles, and embodied intelligence. In just four years, the company has charted a differentiated path that stands in stark contrast to the industry’s trend of “piling on parameters and chasing large models.”

The Beginnings of an Entrepreneurship Journey from a Tsinghua University Lab

Mianbi Intelligence has deep roots in Tsinghua University. The company originated from the Natural Language Processing and Social Humanities Computing Laboratory in Tsinghua University’s Department of Computer Science. When it was officially incorporated in 2022, it was still a micro-enterprise focused on innovation in large-scale model technology. Among the early team members, co-founder and Chief Scientist Liu Zhiyuan is a renowned professor in Tsinghua University’s Department of Computer Science, while Li Dahai, who officially assumed the role of CEO in 2023, previously served as Chief Technology Officer and Executive Director at Zhihu and possesses extensive experience in bringing internet technologies to market.

Liu Zhiyuan (left), Li Dahai (right)

Unlike most large-model startups, which aim for base models with hundreds of billions of parameters from the outset, Mianbi Intelligence laid the groundwork for an “edge-first” approach right from its inception. In 2023, Zhihu led its angel round, in 2024, Huawei’s Habor and Primavera Capital led the Series A round, with Jingguo Rui (a subsidiary of Beijing State-owned Assets), the Beijing Artificial Intelligence Industry Investment Fund, and the Zhongguancun Science City Fund continuing to invest. This Haidian-based startup has gradually earned recognition from both the industry and the investment community. In the first quarter of 2026, China Telecom led the first round of financing, while Shenzhen Venture Capital and Huichuan Industrial Investment—a subsidiary of Huichuan Technology—jointly led the second round. With Beijing and Shenzhen, two major hubs of science and technology innovation, both placing their bets on the company, and Mianbi Intelligence officially joined the ranks of unicorns developing full-stack, self-developed foundation models.

Core Team: Tech-Driven Long-Term Thinkers

Mianbi Intelligence’s core team covers virtually the entire value chain of large language models, from training frameworks to practical applications. Chairman Li Dahai, as the former CTO of Zhihu, understands both the underlying technical logic of large language models and the user pain points involved in bringing internet products to market. In numerous public appearances, he has consistently emphasized that “on-device AI is not about showing off technical prowess, but about bringing intelligence directly to users” devices.” Co-founder and COO Lei Shengtao is primarily responsible for ecosystem partnerships on the industry side, driving the deep integration of model technology with real-world industrial scenarios such as the automotive, manufacturing, and mobile phone sectors. The team led by CFO Luo Lanqian has supported the company’s refined operations following multiple rounds of substantial financing, ensuring that the rapidly expanding enterprise maintains a stable cash flow rhythm.

At the technical team level, the “model scaling technology” proposed by Li Yuxuan, Head of AIInfra at Mianbi Intelligence, became the key to solving the challenge of training accuracy on domestic computing platforms. Rather than blindly following the training approaches of overseas large models, the entire technical team chose to conduct independent R&D starting from the underlying framework, ultimately delivering several industry-first achievements, including the ForgeTrain training framework and the BitCPM three-value large model. Over the course of nearly three years, this team—with an average age of less than 35—transformed the idea of “achieving greater capabilities with fewer parameters” into a technical reality recognized by developers worldwide.

Debunking the Scale Myth: Redefining the “Law of Density” for Large Models”

While the large-model industry generally adheres to the “scale law”—the belief that “the more parameters, the greater the capability”—Mianbi Intelligence has taken the lead in proposing the “density law,” which challenges conventional wisdom: the knowledge density of large models doubles every 3.3 months, meaning that achieving the same level of capability now requires only half the number of parameters previously needed. This theory is not mere speculation but is backed by concrete technological achievements.

The MiniCPM5-1B edge model, released in 2026, achieved a score of 17.9 on the internationally recognized Artificial Analysis benchmark with just 1B parameters—a difference of only 0.4 to 0.7 points from the 200B-parameter GPT-4o released in 2024. achieving the industry’s previously predicted goal of “edge models matching GPT-4”s capabilities by the end of 2026” half a year ahead of schedule. This model was fully pre-trained using the ForgeTrain framework, independently developed by Mianbi Intelligence. It is also the world’s first production-grade large-model pre-training framework written entirely by AI, with a training speed 10% faster than NVIDIA’s Megatron.

In the field of model compression, Mianbi Intelligence is also at the forefront of the industry. They have integrated 1.58-bit ternary quantization-aware training throughout the entire pre-training process, achieving accuracy alignment from the very beginning of training. Their BitCPM-CANN is China’s first ternary large model to be developed entirely on a domestically produced computing platform, capable of end-to-end training, and released as open source. The entire pipeline was natively developed on Huawei’s Ascend platform. During the inference phase, it unlocks approximately a 6-fold increase in GPU memory efficiency while maintaining model capability retention within the range of 90% to 97.2%. The team has now completed in-depth collaborative optimization with Huawei, minimizing the additional overhead of low-bit-width quantization training to an extremely low level. This has achieved an efficiency of 95% relative to standard full-precision training, validating the feasibility of running extremely low-bit-width large models on domestic computing platforms.

In addition to building out its domestic computing power ecosystem, Mianbi Intelligence has also entered into a global strategic partnership with Qualcomm. The collaboration between the two companies has extended to the level of co-designing chips and models, creating optimal integrated hardware-and-software solutions for end devices. At the same time, the team is deeply involved in the development of the FlagOS software ecosystem led by the Beijing Academy of Artificial Intelligence, coordinating chip adaptation efforts across the entire industry through top-level planning to drive the overall improvement of the domestic edge AI computing ecosystem.

Multi-Track Implementation: Comprehensive Penetration from Mobile Phones to Embodied Intelligence

By 2026, the edge AI market will no longer be just a concept confined to PowerPoint presentations. Mianbi Intelligence’s technological achievements have already been implemented at scale across multiple core sectors. In the smart cockpit sector, vehicle models equipped with its production-grade multimodal model can complete the full closed-loop process of “perception–memory–reasoning–execution” without relying on the cloud. Not only can they automatically adjust windows and air conditioning based on the in-car environment; in the event of a vehicle accident, they can also automatically identify the situation, provide emotional support, and guide users through the insurance claims process, thereby integrating the entire accident handling service chain. Taking the Geely Galaxy M9 as an example, the proportion of users who proactively enable on-device AI features far exceeds industry expectations, directly demonstrating users’ recognition of the on-device intelligent experience.

In the consumer electronics sector, Mianbi Intelligence has partnered with Samsung Mobile; the MiniCPM series of on-device models will be integrated into several Samsung flagship models set to hit the market, and its models have also been deployed in products from multiple leading domestic and international smartphone manufacturers. In the specialized terminal sector, Mianbi Intelligence’s models have been integrated into devices such as drones and submersibles, achieving commercial deployment in scenarios including low-altitude flight and civil aviation. At the 2026 Zhongguancun Forum, an embodied robot equipped with MiniCPM-V 4.5 made its public debut. This 8B model is the industry’s first edge model to support high-refresh-rate video understanding, capable of capturing motion details at a rate of 10 frames per second to enable real-time observation, real-time thinking, and real-time response.

To support the developer ecosystem, Mibai Intelligence has launched the Songguo Pi on-device AI development board and the EdgeClaw Box hardware product, establishing a fast-track for rapid validation from algorithms to applications. Currently, the MiniCPM series of open-source models covers all categories of language, full-modal, multimodal, and speech models. Mibai Intelligence is the only vendor in China—apart from Alibaba—to offer a complete suite of open-source on-device models, with cumulative downloads on global platforms such as GitHub and Hugging Face exceeding 24 million.

Industry Landscape and Future Outlook

According to forecasts by global research firm Frost & Sullivan, the global edge AI market is projected to grow from $321.9 billion in 2025 to $1.22 trillion in 2029, at a compound annual growth rate of 40%. China has already incorporated the “smart economy” into its 2026 Government Work Report, setting a target for the penetration rate of next-generation smart devices to exceed 70% by 2027. The dual benefits of policy and market forces are rapidly materializing.

At the 2026 Beijing Zhiyuan Conference, Li Dahai publicly stated that edge-cloud collaboration will be the mainstream development path for edge AI in the future: context management and high-frequency inference tasks will be performed locally at the edge, while the cloud will provide supplementary computing power and capability support, creating a complementary and collaborative relationship between the two. As a number of domestically developed storage-and-computing-integrated edge AI chips have successively completed tape-out, their competitiveness in terms of power consumption, computing power, and bandwidth will gradually become apparent, and the period of concentrated growth for edge applications is not far off.

As Mianbi Intelligence continues to expand its business scope—from its existing ventures in smartphones, automobiles, and embodied intelligence—it will extend into the industrial, home, and specialized sectors in the future. This startup, which emerged from a Tsinghua University laboratory, has not followed the trend of the parameter race in large language models. Instead, it has chosen a “small-is-big” technical approach, using extreme model density to truly bring AI intelligence to every end device. While the industry is still debating where the next breakthrough in large-scale models—reaching 100 billion parameters—will occur, Mianbi Intelligence has already quietly integrated GPT-4-level intelligence into everyone’s smartphones, cars, and robots.