NVIDIA’s New Open 30B hybrid Model Just Beat GPT-OSS and Qwen at 4× Speed!
NVIDIA just did something no one expected from the world’s biggest GPU company – it dropped a genuinely open, long-context frontier-class model family. In this video, we break down Nemotron 3 Nano, the first release in Nvidia’s new Nemotron 3 lineup: a 30B-parameter Mixture-of-Experts model that only activates about 3.6B parameters per token, runs locally on a single 24 GB GPU, and supports up to one million tokens of context without collapsing in speed. We walk through how it delivers 2–3× higher inference throughput than GPT-OSS-20B and Qwen3-30B in real long-generation workloads, why Nvidia ripped out most of the transformer attention stack and replaced it with Mamba-2 state-space layers, and how multi-token prediction, sparsity, and hardware-aligned expert routing make it feel like a completely different beast from normal dense LLMs. We also delve into the aspect that almost nobody is discussing enough: openness. NVIDIA didn’t just drop weights – it released pretrained and post-trained checkpoints, the reward model, NeMo Gym environments, training and fine-tuning recipes, and large chunks of the pretraining and alignment datasets. In a world where Meta is moving close, Chinese models are getting banned from government systems, and OpenAI keeps its data fully opaque, Nvidia is positioning Nemotron as transparent AI infrastructure for governments, enterprises, and serious agentic systems. Then we look ahead to Nemotron 3 Super (~120B) and Ultra (~480B), with latent Mixture-of-Experts, NVFP4 4-bit training, 25T-token corpora, and the possibility that for the first time ever, near-frontier AI might exist in a truly open form – as long as you’re running on Nvidia hardware. By the end, you’ll see why Nemotron 3 Nano isn’t “just another model release,” but a strategic move that could redefine what open-source AI means at the very top end. And if you want the real story behind the world’s fastest-moving tech and AI breakthroughs, make sure to like and subscribe to Evolving AI for daily coverage.