NVIDIA open-sources the Llama Nemotron-253B inference model, with a throughput 4 times higher than that of DeepSeek R1.

NVIDIA announces the open-sourcing of the Llama Nemotron-253B inference model, which is fine-tuned based on Llama-3.1-405B. In multiple benchmark tests, Llama Nemotron outperformed Llama 4, achieving performance comparable to DeepSeek R1 with only half the number of parameters. It excels in complex mathematical reasoning, scientific question answering, and coding tasks, with a throughput that is 4 times higher than that of DeepSeek R1.