NVIDIA recently released Nemotron 3 Ultra — the largest open AI model in the Nemotron 3 lineup. The release took place on June 4, 2026: the model's weights, training data and the training methods themselves were made openly available under a free license. The model is designed for long-running autonomous agents and complex reasoning.
Unlike closed flagships such as ChatGPT or Claude, Nemotron 3 Ultra can be downloaded, fine-tuned on your own data and run on your own infrastructure. The bet here is not on maximum intelligence, but on openness, efficiency and control over the model.
Want more exclusive news and analytics? Subscribe to our Telegram channel, discuss the news and share your opinions about the latest market events in the chat!
What makes the model's architecture special
Nemotron 3 Ultra is not just a «scaled-up transformer». At its core lies a hybrid architecture consisting of three different approaches: Mamba-2 layers, attention layers (Attention) and a latent mixture of experts (Latent MoE) — a mechanism that routes each query only to the relevant «specialists» inside the model.
Mamba-2 layers process long texts quickly and economically: their costs grow linearly with length rather than avalanche-like, as with the usual attention mechanism. Attention layers, in turn, accurately hold large volumes of text in memory. And Latent MoE compresses the data before passing it to the experts, so each of them works narrowly and precisely without requiring additional computation.
In total the model has about 550 billion parameters, but only roughly 55 billion are engaged to process each token. Because of this it thinks like a huge system, while in terms of cost it behaves like a far more compact one. Together with a context window of 1 million tokens and a speed of over 300 tokens per second, this delivers five to six times greater throughput and roughly 30% lower task cost.
Source: BeInCrypto
