The new model is said to be efficient at handling text queries and coding. (Image: NVIDIA)Software giant Nvidia seems to have silently released its latest open-sourced, fine-tuned Large Language Model. Named Llama-3.1-Nemotron-70B-Instruct, the new LLM has reportedly outperformed industry giants like OpenAI’s GPT-4o and Anthropic’s Claude 3.5 Sonnet on some key benchmarks.
The latest LLM is customised by Nvidia and is reportedly useful in terms of LLM-generated responses to general and coding user inquiries. Its advanced architecture and training methodologies have made it lightweight when compared to GPT-4o mini and Meta’s Llama models.
The Llama 3.1 Nemotron-70B model builds on the Llama 3.1 architecture which is based on a transformer technology. It offers 70 billion parameters which allows it to process and generate human-like responses that are coherent and fluent. When it comes to performance, the model has achieved top scores on alignment benchmarks like Arena Hard (85.0), AlpacaEval 2 LC (57.6), and GPT-4-Turbo MT-Bench (8.98).
Based on these scores, the new model surpasses GPT-4o and Claude 3.5 Sonnet across numerous metrics. It needs to be noted that when compared to these models, the new model is significantly smaller with just 70B parameters. NVIDIA has open-sourced the model, reward model, and training dataset on Hugging Face and it can be tested in preview on the company’s official website.
While NVIDIA’s chipmaking feats are known, however, it has been on a spree of producing powerhouse models. The new Nemotron model is a testament to the fact that smaller and more efficient models can compete or even outshine some of the industry leaders.