MOSS Team Startup MOSI Raises Over RMB 1 Billion Pre-A to Build Real-Time Omni-Modal AI for Glasses, Cars and Robots

SHANGHAI — MOSI, the Shanghai startup incubated from Fudan University professor Qiu Xipeng's OpenMOSS team, announced on Monday that it has closed a Pre-A round of more than RMB 1 billion (roughly $140 million), its second disclosed financing of the year. The round was led by Huakong Fund, with participation from CDB Science and Innovation Fund, Shanghai S&T Innovation Fund, Xinwei Capital, Xuhui Capital, Zhuiguang Hard-Tech Ventures and Linggang Ceyuan Fund, while existing backers including Yuanhe Holdings and Shanghai Science and Innovation Group doubled down. Duowei Capital acted as sole financial advisor.

MOSI is betting on a layer that most frontier labs treat as a feature rather than a product: real-time, omni-modal interaction. The company's thesis is that the next wave of AI endpoints will not be chat windows but glasses, toys, cars, home terminals and robots — devices where the model has to continuously hear, see and understand a live environment instead of waiting for a prompt, then turn that understanding into speech, content, tool calls or physical action.

The output so far is unusual for a company at Pre-A stage. In the first half of 2026 alone, MOSI released more than 20 models across speech, vision and real-time multimodal interaction, and kept them open source. The portfolio spans MOSS-TTS for speech generation, MOSS-Audio for audio understanding, MOSS-Transcribe-Diarize for speech recognition with speaker diarization, MOSS-VL for vision-language understanding and MOVA for content generation. In one week in April, MOSS-VL and MOSS-Audio iterated through six versions; in early July the team open-sourced MOSS-Transcribe-Diarize 0.9B and MOSS-VL-Realtime on consecutive days.

The next release, MOSS Live, is the architectural statement. It pairs a front-end interaction model that continuously sees, hears and talks with a back-end agency model that handles complex reasoning and task execution — so an AI can work through a task while staying in conversation with the person in the room. The system targets AI glasses, smart vehicles, home terminals and robotics, where latency and continuous perception decide whether the product is usable at all.

The lineage explains the ambition. In February 2023, Qiu's team released MOSS, the first ChatGPT-style open-source model out of China, and followed with SpeechGPT for end-to-end voice interaction and AnyGPT for unified modeling across text, speech and images. MOSI was spun out of that research at the end of 2024, with CEO Li Shimin coming out of the same group. In March 2026, the company raised a several-hundred-million-RMB angel round led by IDG Capital.

MOSI is also pushing what it calls AI for AI: using generative models to produce speech, image, video and scenario data, then using understanding models to judge, filter and feed results back into the next training round — shrinking the share of the data pipeline that still depends on human annotators. If the approach holds up, the marginal cost of high-quality multimodal training data falls with compute rather than with headcount.

The open question is commercial. Real-time interaction is a crowded lane — every major lab ships a voice mode, and hardware partners have plenty of model vendors to choose from. What MOSI has that most Pre-A startups do not is a research lineage going back to China's first open ChatGPT-style model, a steady open-source cadence and, as of Monday, a war chest sized like a Series B. Whether that converts into design wins in glasses, cars and robots will decide if the RMB 1 billion buys a category or just a nicer research lab.