Ai2's OLMo 2 Challenges Meta's Llama with Fully Open Language Models

Ai2 is a nonprofit AI research organization founded by Microsoft co-founder Paul Allen. The organization has introduced its latest language model, OLMo 2. OLMo stands for “open language mode,” and it stands out as a fully open-source AI tool that meets the Open Source Initiative Criteria.  Unlike other traditional open models, OLMo 2 ensures complete transparency by openly sharing training data, code, checkpoints, recipes, and evaluations. The release of this model supports the goal of overall Ai2 organization in the open source community by offering tools that help to improve AI technology.  The OLMo 2 family includes two models: OLMo 7B with 7 billion parameters and OLMo 13B with 13 billion parameters. Parameters influence problem-solving capabilities and analyze how effectively the AI models perform specific tasks.  The Ai2 organization trained these models on a dataset of 5 trillion tokens from carefully selected high-quality websites, academic papers, Q&A forums, and synthetic and human-generated math workbooks. Ai2 said that “OLMo 2 models are highly competitive with other leading open models like Meta's Llama 3.1”. Notably, the OLMo 2 7B outclass Llama 3.1 8B across a range of tasks, reflecting significant advancements over its success.  Ai2 considers OLMo 2 the most capable fully open-source language model to date. These models, released under the Apache 2.0 license, are available for download from the Ai2 website.  Despite their promise, open-source AI models have sparked debate regarding safety concerns, particularly around their misuse. Reports have indicated that some open models, like Llama, have been repurposed for defense applications by researchers in China.  When asked about the potential misuse of OLMo models, Ai2 engineer Dirk Groeneveld accepted the risks but emphasized the benefits. He argued that “the openness of these models improves ethical advancements, allows verification and reproducibility, and reduces power concentration.