On October 12, MOSS AI announced the completion of a Pre-A funding round exceeding RMB 1 billion, led by Huakong Fund. Major investors included CDB Technology Innovation, Shanghai Science and Technology Innovation Fund, New Alliance Capital, Xuhui Capital, Zhuiguang Hard Tech Venture Capital, and Lingang Ceyuan Fund. Existing investors, including Oriza Holdings, Shanghai Science and Technology Venture Capital Group under Shanghai State-owned Capital Investment, and Xinglian Capital, also participated in the round. Lighthouse Capital served as the exclusive financial advisor for this round and continues to advise the company on a long-term basis.
The funding will support research, development, and capability building in real-time omnimodal interaction intelligence, as well as global commercialization. Following the round, MOSS AI will continue investing in two core directions: real-time omnimodal interaction intelligence and AI for AI. The first addresses how foundation models can operate in real-world settings and continuously understand people and their surroundings. The second brings AI more deeply into data generation, curation, and model iteration itself.
Over the past few years, the fiercest competition in AI has centered on model capabilities: reasoning, coding, and tool use. Foundation models have begun taking on increasingly complex tasks. But as these models become more capable, a new question is emerging: How can this intelligence move beyond the input box and become part of people's lives?
MOSS AI has chosen interaction as its starting point. This reflects its view of how people will access AI in the next phase, while also connecting a decade of research by Qiu Xipeng's team, from pretraining and language models to speech, multimodality, and real-time interaction.
The Next AI Battleground: Beyond the Screen and Toward Real-Time Omnimodal Interaction Intelligence
The first wave of competition among foundation models took place largely on screens. People translated real-world problems into text, speech, or prompts, then asked AI to answer questions, generate content, write code, or use tools.
This approach is already powerful and has shaped much of the vision for AI applications over the past few years. But it rests on one assumption: People must first make sense of reality, then convert it into input the model can process.
AI's next destinations extend beyond the screen. Glasses, toys, cars, home devices, and robots are becoming new points of access. In these settings, users may have no keyboard, and they will not always pause to formulate a complete question. They speak while walking, interrupt themselves, and change their minds. Sounds, images, movements, surroundings, and device states change at the same time. What the real world presents to AI is not a single, neatly prepared input, but an unfolding situation.
As the setting changes, so does the problem AI must solve. The next generation of AI cannot simply wait for a question and return an answer. It needs to hear, see, and understand what is happening, determine what the user truly needs at that moment within a changing context, and respond or act in time. Moving from isolated input-output exchanges to continuously understanding an unfolding world is an essential capability for AI beyond the screen.
The most intense competition has remained concentrated in software: coding, office productivity, and content generation. These applications essentially help models solve problems that have already been defined. To become part of everyday life more broadly, AI must confront a different challenge: reality does not organize tasks for the model in advance.
The industry's growing maturity is also creating the conditions for real-time omnimodal interaction intelligence. Previously, limited model capabilities, high inference costs, and a lack of suitable hardware made it difficult for AI to move beyond screens. Today, open-source AI foundations are maturing, omnimodal inference costs are falling, and multimodal hardware such as AI glasses, AI toys, smart vehicle cockpits, and robots is entering the market. As models, costs, and access points evolve together, bringing AI into the physical world is shifting from a vision to a concrete engineering challenge.
What is truly scarce in this transition is not a model that is better at chatting, but interaction intelligence that enables AI to remain continuously “present.” MOSS AI defines this capability as “real-time omnimodal interaction intelligence”: bringing language, sound, images, environmental information, and device states into a continuous interaction process, then translating understanding into speech, content, tool use, and even device actions.
From MOSS to Real-Time Interaction: Moving Deeper into the Real World
MOSS AI's work on real-time omnimodal interaction is not a sudden change of direction. Viewed across the past decade, Professor Qiu Xipeng's research has moved steadily from language toward people, and then toward the real-world situations in which people live.
Qiu Xipeng, founder and chief scientist of MOSS AI, distinguished professor at Fudan University, and vice dean of the Shanghai Innovation Institute, began researching pretraining in 2016. In February 2023, shortly after ChatGPT brought foundation models into the spotlight, he led his team in releasing MOSS, China's first open-source ChatGPT-like model. GPT-4 arrived a month later, and large language models rapidly became the industry's central focus.
After MOSS, however, the team did not stop at making language models larger. These models had already demonstrated that machines could understand and generate complex language. Qiu soon turned to the next question: Once machines can speak, how can they truly communicate with people?
Human expression always occurs within a specific context. When something is said, the tone in which it is spoken, where the speaker is looking, and what is happening nearby can all change its meaning. If models are ultimately to share our world, they cannot remain limited to understanding neatly prepared text.
The team's laboratory research developed progressively around this question. SpeechGPT, introduced in 2023, explored end-to-end speech interaction. AnyGPT, introduced in 2024, sought to bring text, speech, images, and other modalities into a unified framework. From language to speech and then to multimodality, the research expanded from individual pieces of text to the multiple signals present in actual exchanges between people and AI.
In late 2024, MOSS AI emerged from the OpenMOSS team led by Qiu Xipeng, building on its accumulated research in language, speech, and multimodality. The Shanghai Innovation Institute, where the team is based, also provided talent, computing resources, and an environment for collaboration between academia and industry. From that point, a largely research-driven technical trajectory entered a phase of focused engineering.
A succession of models followed, including MOSS-TTS, MOSS-Audio, MOSS-Transcribe-Diarize, MOSS-VL, and MOVA, covering speech generation, audio understanding, speech recognition, visual understanding, and content generation. In just six months during 2026, MOSS AI released more than 20 models spanning speech, vision, and real-time multimodality, while continuing to make its work open source.
The transition from MOSS to real-time omnimodal interaction involves more than a change in model architecture. The underlying question has evolved from whether machines can understand language to whether they can continuously understand people in a changing real-world situation.
From Models to Systems: AI for AI Connects Interaction with Model Iteration
To continuously understand people and respond promptly in changing situations, AI must coordinate perception, understanding, generation, and execution. MOSS AI's real-time omnimodal interaction system is being built around this chain of capabilities.
Once speech, video, text, and environmental information enter the system, models such as MOSS-VL, MOSS-Audio, and MOSS-Transcribe-Diarize handle understanding. At the other end, MOSS-TTS, MOVA, and models with agentic capabilities handle generation and execution. Between them, the system must manage memory, context, task states, and real-time orchestration across different models.
MOSS Live, which MOSS AI plans to release soon, brings these capabilities together as a system. It pairs a frontend interaction model with a backend agentic model: the former continuously sees, listens, and speaks, while the latter handles complex reasoning and task execution. This allows AI to maintain real-time communication with people while working on their tasks.
“Real time” here means more than faster responses. Information in the physical world carries important temporal relationships. A user may change their mind halfway through a sentence. A new image may alter the meaning of what was just said. Devices and surroundings are also constantly changing. If a model understands the scene as it was several seconds ago, even a correct answer may address the wrong question.
The central challenge of real-time omnimodal interaction is whether AI can keep pace with an unfolding situation.
As systems become more complex, model iteration itself presents a new challenge. Since the beginning of 2026, MOSS AI has released more than 20 models in just over six months across speech understanding, speech generation, visual understanding, and real-time multimodality. During the week of April 7–13 alone, MOSS-VL and MOSS-Audio went through six version releases. In early July, the team released MOSS-Transcribe-Diarize 0.9B and MOSS-VL-Realtime as open-source models on two consecutive days. At this pace, heavy reliance on manual data generation, curation, quality assessment, and feedback would quickly become the next bottleneck.
This is another reason MOSS AI is pursuing AI for AI. It aims to bring AI more deeply into the process of building models themselves: generative models produce new speech, images, videos, and scenario data; understanding models assess, filter, and provide feedback; and the results feed into the next round of training. Many repetitive steps previously performed by people can gradually be handled through collaboration among models.
As these capabilities extend into glasses, cars, home devices, and robots, AI gains access to more than synthetic data. It also encounters real-world data and continuous interactions from different domains, which are relatively scarce in internet datasets. The frontend omnimodal system enters and understands the scene, while the backend AI for AI process feeds new data and feedback into model iteration. The two directions thus become connected.
With this funding round, MOSS AI is investing in a longer, more complete cycle: from models to interaction, from interaction to data, and from data back to models.
Over the past few years, the foundation model industry has focused on how capable models can become. As AI moves beyond the screen, new questions arise: Can it truly understand an unfolding world? Can it interact naturally with people? And can it turn those real interactions into the fuel for its next stage of development?
As early as 2013, Samantha in the film Her offered a vision of an ideal AI assistant. Theodore places his device in his shirt pocket, allowing Samantha to accompany him through streets, see the crowds, and compose a piece of music by the sea for a moment they have shared. Thirteen years later, real-world AI can write code, solve difficult problems, and generate images and videos. Yet we still habitually open a chat window, translate the world before us into text, and wait for an answer. AI knows more and more, but people still have to explain their circumstances to it themselves. If a model can answer complex questions, can it also understand why we are asking this particular question, here and now?
Perhaps understanding begins with sharing the same scene. In Taking Her Eyes With You, Liu Cixin opens up this possibility with a quiet sentence: “I was going on vacation, and the director asked me to take another pair of eyes along.” What one person sees reaches someone far away. Flowers on the grassland and the distant sky become a present moment they can experience together. From the “eyes” taken on a journey to Samantha accompanying a person through the world, these stories express a shared hope: that communication can be grounded in a common context, and that intelligence can understand life as it unfolds. Today, advances in omnimodality and real-time interaction are making that hope technically possible.
This is the direction MOSS AI continues to explore: coordinating perception, understanding, communication, and execution within changing real-world situations, so that intelligence keeps pace with people and with life. MOSS AI hopes that everyone can have their own “Samantha”: an assistant that understands the present situation, picks up unfinished thoughts, respects intentions and authorization, and turns ideas into executable actions and deliverable results. People should be able to receive intelligent assistance without sitting down at a computer for each interaction or putting life on pause.
Connecting AI, humans, and the physical world through a real-time omnimodal interaction system points toward a future found in every journey, every conversation, and every idea that an ordinary person can finally put into action. It brings the most advanced intelligence into everyday lives that are meaningful and worth experiencing.
This is MOSS's vision for the future, and the reality MOSS AI is building.