On October 12, Moss Intelligence (Shanghai Moss Intelligence Technology Co., Ltd.) announced the completion of over 1 billion yuan in Pre-A round financing, led by Huakong Fund, with participation from institutions including Guokai Venture Capital, Shanghai Science and Technology Innovation Fund, Xuhui Capital, and Lingang Ceyuan Fund, while existing shareholders such as Yuanhe Holdings, Shanghai Science and Technology Innovation Group under Shanghai International Group, and Xinglian Capital continued to invest. The company did not disclose its valuation for this round, and the funds will mainly be used for research and development of omni-modal real-time interactive intelligence and global commercialization.

This is about half a year since Moss Intelligence's previous financing round was publicly disclosed. On April 1 this year, the public account of Shanghai Future Industry Fund disclosed that Moss completed a angel round worth hundreds of millions of yuan, led by Shanghai Future Industry Fund and IDG Capital.

Moss Intelligence was founded in November 2024, headquartered in Shanghai, and incubated from Shanghai Innovation Institute and Fudan University's OpenMOSS team. The OpenMOSS official website currently lists Shanghai Innovation Institute, Fudan University's NLP Lab, and Moss Intelligence as joint participants. Founder and Chief Scientist Qiu Xipeng is a Distinguished Professor at Fudan University and Vice Dean of Shanghai Innovation Institute. In February 2023, he led the team to release the open-source large language model MOSS. Co-founder and CEO Li Shimin is a 2022 master's and doctoral continuous student he supervised at Fudan. According to the company's official website, about 50% of the team holds doctoral degrees.

After MOSS, the team's research gradually expanded from text to speech and multimodality. The 2023 SpeechGPT explored discrete end-to-end speech interaction, and the 2024 AnyGPT further attempted to incorporate modalities such as text, speech, and images into a unified discrete token system.

During the same period, foundation models beyond text also saw large capital transactions. On September 28, AMD announced the signing of a final acquisition agreement with spatial intelligence company World Labs, with a transaction value of approximately 8.2 billion USD, expected to be completed by the end of the year, subject to regulatory approvals and other closing conditions. World Labs mainly develops spatial intelligence models that can generate, reconstruct, and simulate interactive 3D worlds.

The two companies have different technological paths, and financing and mergers cannot be directly compared, but World Labs at least provides a contemporaneous reference: foundation model assets beyond text can also attract large capital investments.

According to Moss Intelligence's plan, this round of funds will mainly be directed toward two areas: first, omni-modal real-time interactive intelligence, enabling AI to continuously acquire and understand multimodal information in scenarios such as glasses, cars, and home terminals; second, AI for AI, allowing AI to participate more in data production, filtering, and model iteration.

After its establishment, Moss Intelligence continued the open-source route of the OpenMOSS team. Public repository records show that MOSS-VL released its first batch of Base and Instruct versions in April, launched the real-time video understanding version MOSS-VL-Realtime in July, and subsequently opened quantized versions and technical reports, integrating with open-source toolchains such as SGLang. The speech recognition model MOSS-Transcribe-Diarize currently has over 100,000 downloads on Hugging Face and won first place in the second MLC-SLM Speech Challenge. In July, the Moss API platform began public beta testing.

However, activity in the open-source community and commercial transformation are still two different matters. In this financing disclosure, Moss did not announce named customers, nor did it disclose commercialization data such as revenue, orders, or paying customers. In terms of performance, MOSS-VL-Realtime publicly disclosed a score of 66.0 in the PA@OmniMMI test and has released a technical report, but in currently public information, there is no direct horizontal comparison with mainstream real-time multimodal models such as GPT-4o and Gemini under unified testing conditions.

Several state-owned background institutions also appeared in this round's investment list. Shanghai Science and Technology Innovation Fund, Xuhui Capital, and Lingang Ceyuan Fund all have Shanghai local state-owned backgrounds, while Shanghai Science and Technology Innovation Group under Shanghai International Group continued to invest. The company did not disclose this round's valuation, the corresponding equity ratio of the financing, or board arrangements, and the specific valuation cannot be inferred solely from the financing amount.

In the financing announcement, Moss Intelligence also previewed that the real-time interactive system MOSS Live is about to be released, working collaboratively with a front-end interaction model and a back-end agentic model. For this company less than two years old, the more verifiable indicators going forward are the actual interaction performance after MOSS Live's release and whether the API platform can generate paying customers and revenue.