Mosi Intelligence closed a Pre-A round of more than RMB 1 billion, in news that surfaced on October 12, 2026. According to The Paper's report that day, the round was led by Hua Kong Fund, with participation from CDB Sci-tech, Shanghai Sci-tech Fund, Xinwei Capital, Xuhui Capital, Zhuiguang Hard-tech Ventures and Lingang Ceyuan Fund, while existing backers including Yuanhu Holdings, Shanghai Sci-tech Group under Shanghai Guotou, and Xinglian Capital followed on. Duowei Capital acted as exclusive financial advisor. The company said the money will fund R&D for full-modal real-time interactive intelligence and its global commercial rollout.
State-backed capital fills much of the roster
Read closely, the investor list leans heavily on state-linked science funds: the lead investor plus CDB Sci-tech, Shanghai Sci-tech Fund and Lingang Ceyuan Fund account for a large share. For a company incubated only at the end of 2024, raising more than RMB 1 billion at Pre-A means capital is betting on the team's track record and technical direction rather than current revenue. It is the same pattern as the recent early-stage bet on Tian Keyu's world-model startup: money arrives for people and roadmap before the product matures. The difference is that Mosi already has a stream of released models, not just a plan on paper.
From the OpenMOSS team to a company
Mosi was incubated from the OpenMOSS team at Fudan University, led by Qiu Xipeng, a distinguished professor at Fudan and vice dean of the Shanghai Academy of Artificial Intelligence. The team has a long record in language, speech and multimodal research, and its MOSS series was well known in the early phase of China's large-model wave. After spinning out, the team released more than 20 models across speech, vision and real-time multimodal work in the first half of 2026, and kept open-sourcing them. Academic spinouts often struggle with engineering pace and delivery; judging by release cadence in the first half of the year, Mosi has at least switched to a company rhythm on output.
Two bets: real-time interaction and AI for AI
The company has named two clear directions for the money. The first is full-modal real-time interactive intelligence: models that keep understanding people and their surroundings in live settings, listening, watching and responding continuously instead of in turn-based exchanges. Its speech and vision heritage points straight at this, and the hard parts are latency, interruption handling and long-term memory — exactly where it can separate itself from general chat models. The second bet is AI for AI: putting AI into data production, data screening and model iteration itself, using models to improve models. If that works, it lowers the training cost of every future model generation; if it does not, it remains a financing narrative. Both routes point at the same goal: turning research strength into deliverable, billable product capability.
What counts next is delivery, not release count
With more than RMB 1 billion in the bank, expectations shift from how many models Mosi ships to who uses them and who pays. Real-time interaction usually lands first in customer service, in-car systems, smart hardware and companion products — scenarios extremely sensitive to latency and stability, where a polished demo and a production system are separated by a whole engineering stack. Global commercialization also demands local delivery and compliance work. The next real checkpoint for judging this company is not another funding headline, but when its first large-scale paid deployment shows up.
