智东西（公众号：zhidxcom）
作者 | 江宇
编辑 | 李水青
智东西10月12日报道，今日，上海大模型公司模思智能（MOSI）宣布完成超10亿元Pre-A轮融资，由华控基金领投，主要投资方包括国开科创、上海科创基金、新微资本、徐汇资本、追光硬科技创投、临港策源基金等，元禾控股、上海国投旗下上海科创集团、星连资本等老股东持续投资。多维资本担任本轮及长期独家财务顾问。
这是模思智能今年第二次官宣融资。今年3月，模思智能完成数亿元天使轮融资，由IDG资本领投，元禾控股、上海国投旗下上海科创及上海未来产业基金、奇绩创坛、智谱系基金星连资本及某头部产业投资方联合投资。
据官方披露，本轮融资资金将主要用于全模态实时交互智能的研发与能力建设，以及全球商业化落地。
模思智能孵化于复旦大学教授邱锡鹏带领的OpenMOSS团队。2023年2月，邱锡鹏团队推出国内首个类ChatGPT开源模型MOSS。此后，该团队陆续推出语音、视觉及多模态系列模型。
▲邱锡鹏
与智谱、DeepSeek等在通用大模型、推理及Coding等领域持续投入的厂商相比，模思智能更加聚焦语音、视觉及实时多模态交互，让AI能够持续感知和理解真实环境，并应用于AI眼镜、智能汽车、机器人等终端场景。
此外，模思智能还在探索AI for AI，利用AI参与数据生产、筛选和模型迭代。
一、半年推出20多个模型，瞄准全模态实时交互
2026年以来，模思智能进一步加快模型迭代，在半年多时间里发布了20多个模型，并持续推进开源。
其已陆续推出MOSS-TTS、MOSS-Audio、MOSS-Transcribe-Diarize、MOSS-VL、MOVA等系列模型，覆盖语音生成、音频理解、语音识别、视觉理解及内容生成等领域。
其中，仅4月7日至13日一周，MOSS-VL和MOSS-Audio就连续迭代出6个版本。7月初，该团队又连续两天开源发布MOSS-Transcribe-Diarize 0.9B和MOSS-VL-Realtime两款模型。
在技术路线方面，模思智能目前重点布局全模态实时交互，希望让AI能够在持续变化的真实场景中，同时处理语音、画面、环境及设备状态等信息，并根据用户需求及时反馈。
据了解，模思智能即将推出全模态实时交互系统MOSS Live。
该系统采用前端交互模型与后端能动性模型协同工作的架构。前端模型负责持续感知视觉、语音等信息并进行实时交流，后端模型则处理复杂推理和任务执行，使AI在执行任务的同时保持与用户的交互。
这套系统将面向AI眼镜、智能汽车、家庭终端及机器人等应用场景。
除实时交互外，模思智能还在推进AI for AI技术路线。
其核心思路是利用生成模型生产语音、图像、视频等数据，再通过理解模型进行质量判断、筛选和反馈，将结果用于后续模型训练，减少数据生产和模型迭代过程中对人工操作的依赖。
二、复旦教授邱锡鹏创办，CEO出自其研究团队
模思智能创始人兼首席科学家邱锡鹏是复旦大学教授、国家杰出青年科学基金获得者，同时担任上海创智学院副院长、全职导师。
邱锡鹏长期从事自然语言处理与大模型研究，其著作《神经网络与深度学习》被广泛用于高校教学，并多次入选全球前2%顶尖科学家榜单。
2024年，邱锡鹏获得CCF-ACM AI Award（中国计算机学会与美国计算机学会人工智能奖），为当年该奖项全球唯一获奖者。
在成立公司之前，邱锡鹏团队已经开展了多年的大模型研究。
2016年，邱锡鹏开始研究预训练技术。2023年2月，团队发布类ChatGPT开源模型MOSS。随后，团队于2023年推出端到端语音交互模型SpeechGPT，并在2024年通过AnyGPT探索文本、语音、图像等不同模态的统一建模。
模思智能CEO李世民同样出自邱锡鹏研究团队，曾在其指导下硕博连读。李世民以第一作者身份发表过多篇CCF-A类论文，并深度参与科技部“新一代人工智能2030”重点研发项目，研究方向涉及拟人化人机交互等技术。
▲李世民
结语：模思智能加码实时交互，商业化落地待观察
从MOSS语言模型到语音、视觉及实时多模态模型，邱锡鹏团队近年来持续扩展技术布局，相关研究也逐步进入商业化阶段。
目前，模思智能已形成覆盖语音、视觉及多模态交互的模型产品矩阵，即将推出的MOSS Live则将进一步整合这些能力。
后续，其能否在AI眼镜、汽车、机器人等实际场景中实现规模化应用，有待观察。
本文由 AI图鉴自动抓取并翻译，版权归原作者所有；如需查看原文请点下方链接。
Intellectual (WeChat Official Account:zhidxcom)
Author: Zheng Zheng, Jiang Woo
Editor, Lee Shui-qing
On October 12th, Shigeki reported that today, Shanghai’s LLM company, MOSI, announced the completion of over one billion yuan in Pre-A round of financing, which was carried out by the China-controlled fund, with the main investors being the old shareholders, such as Zakai, the Shanghai Foundation, the Shanghai Foundation, the New Micro-capital, Xu Seok, the Light and Hard Technologies Foundation, the Port-of-Hang Fund, and others. Multi-dimensional capital serves as sole financial adviser for the current round and for the long term.
This is the second official announcement this year of the funding of myths. In March this year, Modified Smart completed the hundreds of millions of dollars of angel round finance, with IDG capital invested in a joint venture of dollar-share holdings, Shanghai-based Industrial Fund for the Future of Shanghai and Shanghai-based Industry Fund, the Tribune Forum, Zhipu Foundation Capital and some head-line industry investors.
According to official disclosures, the current round of financing will be used mainly for research and development and capacity-building for all-modular real-time interactive intelligence, as well as for global commercialization.
Momentum Smart hatched in the OpenMOSS team led by Professor Qui Xin Peng at the University of Jordan. In February 2023, the Quixion team launched the country ' s first ChatGPT open source model, MOSS. Since then, the team has introduced voice, visual and multi-modular series models.
Quixion.
Modelling intelligence is more focused on voice, visual, and real-time multimodular interaction than the persistent input of manufacturers such as Zhipu, DeepSeek and others in areas such as General LLM, reasoning and Coding, which allows AI to continuously sense and understand the real environment and apply to terminal scenarios such as AI glasses, smart cars and robots.
In addition, Modular Intelligence is exploring AI for AI, using AI for data production, screening and modelling.
I. Half-yearly roll-out of more than 20 models targeting all-modular real-time interaction
Since 2026, modulation intelligence has further accelerated the iterativeization of models, with more than 20 models of publish over a period of more than six months, and the continuous advance of open sources.
The MOSS-TTS, MOSS-Audio, MOSS-Transcribe-Diarize, MOSS-VL, MOVA and others have been launched, covering areas such as voice generation, audio understanding, voice recognition, visual understanding and content generation.
Of these, during the week of 7 to 13 April alone, six consecutive versions were produced by Moss-VL and Moss-Audio. At the beginning of July, the team was able to produce two more models, publishMOSS-Transcribe-Diarize 0.9B and MOSS-VL-Realtime, for two consecutive days.
In terms of technical routes, the current focus of the modem is to interact in real time, with the hope that AI will be able to process voice, image, environment and device status information in the context of the real scene of continuous change and provide timely feedback based on user needs.
It is understood that Motion Smart is about to launch a full-modular real-time interactive system, MOSS Live.
The system uses the architecture of a front-end interactive model working in tandem with a back-end dynamical model. The front-end model is responsible for the continuous perception and real-time exchange of information such as visual, voice, etc., while the back-end model deals with complex reasoning and task execution, allowing AI to remain interactive with users while performing its tasks.
The system will be aimed at applications such as Ai glasses, smart cars, home terminals and robotics.
Aside from real-time interaction, a memorization intelligence is driving the AI for AI technology route.
The core idea is to produce data such as voice, images, videos, etc., using the generation model, and to interpret the model for quality judgement, screening and feedback, using the results for subsequent model training and reducing reliance on manual processes in data production and modelling over time.
ii. Professor Qui Xin Peng was founded by CEO from his research team.
The founder and chief scientist of the Moot was a professor at the University of Jordan and a recipient of the National Fund for Outstanding Youth Science, as well as a full-time mentor and vice-president of the Academy of Zaiqi in Shanghai.
Qui Qui Qui Peng has long been involved in natural language processing and LLM research, and his book Neurological Networks and Deep Learning has been widely used in higher education and has been selected for the top 2 per cent of the world ' s top scientists on several occasions.
In 2024, Quixion was awarded the CFC-ACM AI Award Award for Artificial Intelligence by the Chinese Computer Society and the American Computer Institute, the only global recipient of the Award for the year.
Prior to the establishment of the company, the LLM study had been carried out by the Qui Sipen team for many years.
In 2016, Quixion began researching pre-training techniques. Team publish ChatGPT open source model MOSS, February 2023. Subsequently, the team launched the end-to-end voice interactive model SpeechGPT in 2023 and explored uniform modelling of text, voice, image, etc. through AnyGPT in 2024.
The smart CEO Lee Shi-min also came from the Qui Si-Peng research team and read it under his guidance. Li Shimin, in his capacity as the first author, has published numerous articles of the CCF-A category and is deeply involved in the Ministry of Science and Technology's “New Generation of Artificial Intelligence 2030” priority development project, which is directed towards technologies such as humanization.
Lee Se-min.
Concluding remarks: Think smart plus code interactive in real time, commercialization pending observation
From the MOSS language model to the voice, visual and real-time multi-modular model, the Quision team has been continuously expanding its technical layout in recent years and its research has gradually entered the commercial stage.
At present, a matrix of model products covering voice, visual and multimodular interactions has been formed, and the forthcoming MOSS Live will further integrate these capabilities.
Following this, it remains to be seen whether it will be possible to achieve a scaleed application in practical scenarios such as AI glasses, cars, robots, etc.
Auto-fetched and translated by AI Atlas. All rights belong to the original author.