Mosi Intelligence raises more than $149M for real-time AI

Mosi Intelligence has closed a Pre-A round of more than $149 million to develop AI systems that can watch, listen and respond as events unfold, rather than waiting for a neatly written prompt. The startup is targeting smart-cockpit and other connected-device makers. Huakong Fund led the round, according to the company, which said existing investors also participated. Figures are in US dollars, converted at October 12, 2026 reference rates.

Its planned MOSS Live system is supposed to combine continuous audio and video with environmental and device data. One interaction model would handle the live conversation, while an agent model would perform longer reasoning, call tools and execute tasks. That architecture is still being assembled from separate models, not offered as a finished general assistant.

The customer problem is concrete in a car: hearing someone say they are cold isn't enough. Software must determine who spoke, whether the person was addressing the assistant, what the cabin is doing and whether the remark was a request to change the temperature. Mosi told 36Kr's Waves that it is developing with enterprise users in smart cockpits and AI hardware, but the report did not name customers.

Mosi's current components cover speech generation and recognition, audio understanding, visual understanding and content generation. Chief Executive Li Shimin told Waves the team had released more than 20 model versions over the previous six months. The point isn't simply to add more input types; it is to keep listening and seeing while responding, then learn from signals such as interruption, silence or frustration that don't arrive as explicit feedback.

Fudan professor Qiu Xipeng founded Mosi after his research group released the MOSS language model in 2023, followed by SpeechGPT and the multimodal AnyGPT. Li, a core SpeechGPT author, became chief executive when the startup was formed at the end of 2024. Mosi said the new capital will support real-time multimodal research and global commercialization; Li also acknowledged that continuous audio-video understanding, long-term memory and ambiguous real-world feedback remain unsolved.