NVIDIA launches Nemotron 3: Open AI models for everyone
NVIDIA just dropped its new Nemotron 3 family—three open AI models built for everything from quick tasks to deep reasoning.
The lineup includes Nano (about 30B parameters), Super (100B), and Ultra (a massive 500B).
The Nano model is especially speedy, promising up to four times more output than the previous Nemotron 2 Nano.
What's inside?
Nemotron 3 uses a smart hybrid Mamba-Transformer "Mixture-of-Experts" setup, which keeps things fast and efficient without losing accuracy.
Nano is great for affordable jobs like debugging or summarizing, while Super and Ultra handle bigger challenges like multi-agent teamwork or complex problem-solving.
Features like LatentMoE and Multi-Token Prediction, available in the Super and Ultra models, help them run even smoother on NVIDIA's latest Blackwell hardware.
Why does it matter?
In benchmarks, Nemotron 3 outperformed rivals like Qwen3 and GPT-OSS-20B on tasks that test long-form reasoning and tool use.
Plus, since these models are fully open, developers can tweak them however they want—making it easier for everyone to build cool new AI tools without barriers.