Key Takeaways:
MiniMax's H3 open-weight model has become one of the fastest-adopted AI releases on Hugging Face, generating nearly 300 community derivatives and roughly 7 million ComfyUI downloads in its first week.
MiniMax's open-source H3 model generated nearly 300 community derivatives and about 7 million ComfyUI downloads in its first week, making it one of the most-adopted open-weight AI models globally. The 33-billion-parameter model, which generates video with natively synchronized stereo audio from text, images, video or audio, topped Hugging Face's Trending Models list.
"Chinese AI models are gaining traction among global developers because of their combination of strong performance, lower costs and greater openness," Xiang Ligang, director-general of the Zhongguancun Modern Information Consumer Application Industry Technology Alliance, said.
The open portion of H3 contains about 60 billion parameters, with BF16 weights requiring roughly 120GB of memory — far exceeding the VRAM capacity of most consumer-grade GPUs. Following the open-source release, the community actively optimized inference efficiency and hardware compatibility. A research team from NVIDIA used Sol Engine on the first day of release to complete the first round of inference optimization within 4.5 hours.
MINIMAX-W shares rose 5.54% on the news, though short selling of $240.05 million at a 6.101% ratio suggests lingering bearish positions. The model's rapid adoption comes as Chinese AI models swept the top four spots on OpenRouter's weekly LLM rankings, with DeepSeek V4 Flash topping the list at 9.39 trillion tokens in usage.
The 120GB memory requirement for BF16 weights initially limited H3 to high-end workstations and data center GPUs. Community-driven quantization efforts and inference optimizations have since expanded the model's reach to more developer devices. NVIDIA's Sol Engine optimization, completed in 4.5 hours on day one, demonstrated how quickly the open-source community can adapt models for broader hardware compatibility. The NVFP4 checkpoint format, optimized for NVIDIA Blackwell GPUs, further reduces memory requirements without sacrificing accuracy.
The model's multimodal capabilities — generating video with synchronized audio from mixed inputs — position it against rivals including DeepSeek's V4 Flash, which recorded 9.39 trillion tokens in weekly usage on OpenRouter, and Zhipu AI's GLM 5.2. Chinese AI models accounted for 34.25 trillion tokens in usage from August 3 through Sunday, up 21.76 percent from the prior week, while US models recorded 9.17 trillion tokens. Tencent's Hunyuan and Xiaomi's MiMo-V2.5 also ranked in OpenRouter's top five, showing the breadth of Chinese open-weight offerings.
The H3 adoption pattern reflects a broader shift in the global AI market. According to UBS estimates, leading Chinese AI models cost about one-tenth as much to train as comparable overseas models, while their API prices typically run 10 percent to 20 percent of foreign alternatives. This cost advantage, combined with open-weight distribution, has driven Chinese models to consistently account for more than 30 percent of weekly token usage by US companies, with the share peaking at 46 percent, according to data from a third-party industry platform cited by People's Daily.
The open-source approach contrasts sharply with many large US AI companies that primarily offer flagship models through closed platforms and APIs. As more developers build on Chinese open-weight models, they create tools and applications that further expand the reach of these models. China's recently released Action Plan for Global AI Governance explicitly encourages the joint development of international open-source AI communities and the sharing of general-purpose large models, foundational algorithms and software tools.
For investors, MINIMAX-W's stock reaction — up 5.54% with $240.05 million in short selling at a 6.101% ratio — reflects both enthusiasm for the model's traction and persistent bearish positioning. The company's ability to convert open-source adoption into commercial revenue will determine whether the current valuation holds. Chinese AI models' growing global adoption also raises questions about the effectiveness of US export controls, as technological restrictions have not slowed innovation in China's AI sector. The US government has reportedly considered banning cutting-edge Chinese AI models, a move that analysts say could backfire by raising costs for US companies and limiting their access to competitive technologies.
This article is for informational purposes only and does not constitute investment advice.
