Ai2, a nonprofit AI research group founded by Microsoft co-founder Paul Allen, has released OLMo 2, its latest family of open-source language models. This release builds on the success of the first OLMo series introduced earlier in 2024.
OLMo 2 comes in two versions: OLMo 7B, with 7 billion parameters, and OLMo 13B, with 13 billion parameters.
Parameters in AI models refer to components that help the model perform tasks like answering questions, summarizing texts, and writing code.
Ai2 says OLMo 2 is competitive with other open models, such as Meta’s Llama series, and even outperforms Llama 3.1 8B in some areas.
The models were trained using 5 trillion tokens, which included data from high-quality websites, academic papers, discussion boards, and both human and synthetic math problems. Ai2 openly shared all data, training code, and checkpoints to ensure transparency and reproducibility.
Ai2 released OLMo 2 under the Apache 2.0 license, making it available for commercial use. The organization emphasizes that making models fully open can promote innovation and equitable access. However, there are concerns about misuse, as seen with other open models used in unintended ways.
Despite risks, Ai2 believes the benefits of open models outweigh potential harms. The full OLMo 2 models and resources are available for download on the Ai2 website. This step strengthens Ai2’s position as a leader in open-source AI.
