As Chinese AI companies expand their influence in the global AI market by promoting open-weight models, NVIDIA and Meta have announced or revealed plans to release open models, creating a new competitive landscape in an AI market that was previously dominated by closed frontier models.
NVIDIA unveiled its 30.9 billion-parameter Mixture of Experts (MoE) model, “Nemo Triton 3.5 Lightning,” on the 11th (local time). Enterprises and developers can download and use it without separate licensing fees or prior approval, and the weights for model tuning are also publicly available, allowing users to freely modify them.
NVIDIA highlighted that, using a so-called “distillation” technique, the model achieves token generation speeds up to 4 times faster than comparable open models and reduces overall agent task completion time by approximately 30%. The company is also reportedly developing its next-generation AI model, “Nemo Triton 4,” with 1 trillion parameters.
Analysts suggest that NVIDIA's advocacy and active development of open AI models are driven by synergies with its AI semiconductor business.
Jensen Huang, NVIDIA's CEO, recently stated in an interview with Axios, “Free AI is inherently good for hardware, semiconductors, and data centers.”
A day earlier, Meta also declared its return to the open ecosystem by releasing an open AI model optimized for personal devices.
In his essay “The Future Belongs to Everyone,” published on the 10th, Mark Zuckerberg, Meta's CEO, opposed the concentration of AI capabilities in a few companies, arguing, “Rather than centralizing superintelligence, it should be widely distributed so that everyone can control it.”
To align with this vision, Meta introduced its next-generation open AI model, “Muse Glimmer,” a compact 30.9 billion-parameter model specialized for on-device AI. It can run agent tasks continuously on a single GPU within a standard PC.
While Meta previously led the open model ecosystem with “Llama,” it shifted focus to closed models after restructuring its AI business last year. This return to an open strategy is interpreted as a move to gain competitiveness in the closed AI market dominated by OpenAI and Anthropic, while also responding to the open-weight offensive by Chinese companies.
In China, major AI firms such as MoonShot AI's “KimiK3,” Alibaba's Qwen, Z.ai's GLM, DeepSeek, and MiniMax have successively released open-weight models, intensifying competition to attract developers. As Chinese models narrow the performance gap with U.S. closed models, relying solely on frontier models has become insufficient for dominating the global AI ecosystem.
Closed AI models excel in achieving state-of-the-art performance through massive computing resources and proprietary technology, whereas open-weight models spread rapidly due to their flexibility, allowing developers and enterprises to modify and deploy them freely.
