NVIDIA is releasing not only model weights but also datasets and training recipes in an open format for its internally developed "Nemotron" series of open models. At GTC in March 2026, the company established the "Nemotron Coalition," involving participants such as Mistral AI and Perplexity, to build a framework for the joint development of frontier models.
The background of this strategy is the expansion of demand for infrastructure such as GPUs. While NVIDIA's revenue is currently centered on sales to a few giant cloud providers, the company is promoting the "AI Factory" concept, where enterprises maintain their own clusters of GPU servers. NVIDIA believes that improving the performance of open models and expanding the user base will ultimately lead to the construction of on-premises environments and the purchase of GPUs by corporations.
Meanwhile, Alibaba Cloud is also aiming for adoption into its own cloud ecosystem by releasing the Qwen series of models for free. Furthermore, the nature of open models is deeply tied not only to corporate strategy but also to competition between nations and geopolitical contexts, such as the movement by Chinese entities to promote openness in response to U.S. export restrictions, and the U.S. government's stated policy to support open-weight models.
The "NVIDIA-Nemotron-3.5-Lightning-30B-A3B-NVFP4" released by NVIDIA is a model trained on more than 20 trillion tokens and supports multiple languages, including English and Japanese. This model is optimized for NVIDIA GPUs and incorporates various technologies to achieve high-speed inference.
Source: