The Allen Institute for AI (Ai2), a nonprofit research organization established by Microsoft co-founder Paul Allen, has launched a new family of open-source language models called OLMo 2. This marks the second installment in the OLMo series, short for "Open Language Model."
Unlike many models labeled as “open,” OLMo 2 aligns with the Open Source Initiative's (OSI) recently established definition of open-source AI. This means all tools, data, and processes involved in its development are publicly accessible, ensuring full reproducibility. Ai2’s earlier models, released in February, also adhered to this standard.
According to Ai2, OLMo 2 was developed with openly shared data, training code, transparent evaluations, intermediate checkpoints, and reproducible training methodologies. The aim is to foster innovation within the open-source community by providing accessible resources and insights.
The OLMo 2 family includes two models: one with 7 billion parameters (OLMo 7B) and another with 13 billion parameters (OLMo 13B). In AI models, the number of parameters correlates with complexity and performance, making these models competitive with leading alternatives such as Meta's Llama 3.1.
Ai2 trained the OLMo 2 models using a dataset containing 5 trillion tokens. These tokens were derived from diverse sources, including carefully curated websites, academic papers, Q&A forums, and math workbooks—both human-generated and synthetic.
The models are designed to handle a variety of text-based tasks, such as answering questions, summarizing information, and generating code. Ai2 asserts that the new models outperform previous OLMo iterations and even rival Meta’s Llama 3.1 in certain benchmarks. The models and their components are available for commercial use under the Apache 2.0 license, which ensures accessibility and flexibility for developers.
While open-source models like OLMo 2 offer transparency and encourage innovation, they also raise concerns about potential misuse. Open models have been controversially employed in areas such as defense tool development. Addressing these concerns, Ai2 engineer Dirk Groeneveld has emphasized the broader benefits of open models, such as fostering ethical advancements, enabling reproducibility, and reducing concentrated control in the AI industry.
“Open models may be misused, but they are essential for equitable access, verification, and advancing ethical practices,” Groeneveld stated.
By releasing OLMo 2, Ai2 continues to champion open-source development, striving to create tools that balance accessibility, innovation, and responsibility.
