中国大模型第三股要来了？

Core Viewpoint - AI model company Jumpspace is considering an IPO on the Hong Kong Stock Exchange within the year, aiming to raise approximately $500 million [1]. Group 1: Company Overview - Jumpspace was established in April 2023 and is headquartered in Xuhui, Shanghai, focusing on achieving AGI (Artificial General Intelligence) and developing foundational models along with AI+ terminal applications [3]. - The company has developed a core matrix of 1+2 in model research, which includes continuous exploration of intelligent limits and significant investment in foundational language models, having released three generations of foundational models [3]. Group 2: Market Context - Other AI model startups like Zhipu and MiniMax have successfully listed on the Hong Kong market, demonstrating strong market performance and providing valuation references for companies in the primary market [2]. - Jumpspace is viewed as a strong candidate for being the third major player in China's large model sector, alongside Kimi, which has no immediate plans for an IPO [3]. Group 3: Recent Developments - Jumpspace completed a B+ round of financing exceeding 5 billion RMB, setting a record for the largest single financing in the Chinese large model sector over the past 12 months [3]. - The financing round included participation from various institutional investors and will be used for foundational model research and to accelerate the AI+ terminal strategy [3]. Group 4: Leadership and Innovations - Jumpspace announced the appointment of Yin Qi as the chairman, who brings extensive experience in AI technology evolution, business judgment, and organizational development [4]. - The company released a new generation of open-source agent foundational model, Step 3.5 Flash, designed for real-time agent workflow scenarios, achieving a maximum inference speed of 350 tokens per second [4]. - Step 3.5 Flash utilizes a sparse MoE architecture, activating approximately 11 billion parameters per token out of a total of 196 billion parameters, significantly enhancing inference efficiency while maintaining model capability [4].