Nvidia has launched Llama-3.1-Nemotron-Ultra-253B, a fully open-source language model that outperforms the larger DeepSeek R1 on several benchmarks despite having less than half the parameters. Carl Franzen of VentureBeat reports the model is now available on Hugging Face with open weights and training data. The 253-billion parameter model features a unique toggle for “reasoning on” and “reasoning off” modes, allowing developers to switch between complex reasoning tasks and straightforward outputs. Performance improves dramatically with reasoning enabled—MATH500 scores jump from 80.40% to 97.00%. The model runs efficiently on a single 8x H100 GPU node, supports sequences up to 128,000 tokens, and works in multiple languages including English, German, French, and Spanish. It’s licensed for commercial use under the Nvidia Open Model License.