OLMo 2 models rival Meta’s best in performance
In a push for transparency and accessibility, Ai2 has unveiled OLMo 2, a powerful new series of open-source language models. Designed to compete with industry leaders like Meta, the models aim to advance AI innovation through open access.
Ai2, a nonprofit AI research group, has introduced OLMo 2, a groundbreaking series of open-source language models designed for transparency and reproducibility. The models, developed using open-access data and tools, align with the Open Source Initiative’s standards for AI, setting them apart from many competitors.
The OLMo 2 series includes two versions: one with 7 billion parameters and another with 13 billion, making them powerful tools for tasks like summarising documents, answering questions, and generating code. Trained on a dataset of 5 trillion tokens sourced from websites, academic papers, and other vetted materials, the models perform competitively against Meta’s Llama 3.1.
While some critics voice concerns about potential misuse of open models, Ai2 argues their benefits outweigh the risks. By making the models freely available under an Apache 2.0 license, the organisation hopes to democratise AI development and promote ethical innovation.
