NVIDIA has announced the introduction of the Llama Nemotron Reasoning model, the LLM model that enhances answer accuracy by 20% through a pre-response thinking approach. By training on data demonstrating thinking methods, NVIDIA has developed a pre-response thinking dataset comprising 60,000 million tokens, involving human assistance for data review which took a total of 45,000 hours and model runtime on H100 for a total of 360,000 hours.
The model comes in three sizes – Nano for ultra-small scale, Super for single server GPU operation, and Ultra for multi-GPU server operation. NVIDIA not only trained the model but also pruned it to ensure faster model runtime. The model is trained to stop thinking if users require quick answers, making it suitable for both normal and pre-response thinking usage similar to Claude 3.7.
NVIDIA is offering model downloads via NIM Micro Services, currently available for Nano and Super versions, with the Ultra version to be released later along with the accompanying thinking methodology dataset. The usage of the models for testing, research, and development is free of charge.
TLDR: NVIDIA introduces the Llama Nemotron Reasoning model, enhancing answer accuracy by 20% with a pre-response thinking approach and offering free model downloads for testing and research purposes.
Leave a Comment