The Future of GenSpark and Flowith

Originally published in Chinese on April 13, 2025. GenSpark and Flowith are two Chinese-founded general agent products that followed Manus. The news below was compiled with Tencent Yuanbao, Tencent's AI assistant, and states facts only. It does not represent my views. Latest on GenSpark Explosive growth and funding (April 2025). GenSpark launched its general AI agent "Super Agent" in April 2025. Within 9 days of launch it hit $10 million in annual recurring revenue (ARR), and it raised $100 million at a $530 million valuation, making it one of the fastest-growing companies in the AI sector. Its users grew from 1 million in November 2024 to 2 million in February 2025, and the iOS app passed 10 million users within three months. A free tier plus a paid subscription ($24.99 a month) drives the commercialization. Technical breakthroughs and feature upgrades. A mixture-of-agents architecture (MoA) integrates 8 large language models of different sizes (such as GPT-4o and Claude 3.7), a toolset of more than 80 tools (a phone-call module, video generation tools, and others), and curated datasets, allocating tasks dynamically and lowering error rates (on the GAIA test its error rate is 20% below the industry average). It supports voice calls (for example, placing an international call to book a restaurant), video generation (a travel video with sound from a stated need), and automatic PPT generation, closing the loop on complex tasks. Market positioning and competitive strategy. It turned from an AI search engine into a general agent, aimed directly at Manus, and differentiates with a lower error rate, real-time interaction, and the phone-call feature. It has a commercial scenario already validated: running a Xiaohongshu account through an AI agent gained 5,000 followers in two weeks and turned a profit, and it plans to tie in deeply with e-commerce and marketing scenarios. Latest on Flowith Feature iterations and market response (April 2025). Flowith launched Oracle mode, which supports breaking complex tasks down and multi-step reasoning. When a user enters "learn to paint," for example, it splits the request into sub-tasks such as choosing a style and basic training, and calls several models to generate a personalized learning plan. Its canvas-style interface (similar to Figma) and multi-person collaboration drew attention: users can run multiple threads at once, such as producing marketing copy and debugging code, and share progress in real time. Core technology and features. Multi-model parallel processing: built on the Oracle engine, it integrates models from OpenAI, Anthropic, DeepSeek, and others