When you buy through affiliate links in our content, we may earn a commission at no extra cost to you. Learn 
The non-profit startup Ai2 has released OLMo 2, a new family of large language models that are fully open source and are claimed to be highly competitive with other top models.
Earlier this week, Ai2 unveiled OLMo 2 – the second generation of their “Open Language Model” family. And when they say open, they really mean it. OLMo 2 meets the criteria recently set by the Open Source Initiative for what qualifies as legitimate open source AI.
That means the training data, code, model weights, and basically everything that went into creating OLMo 2 is available for anyone to inspect, tweak, and build upon. It’s a step up from Meta’s LLama models which, while open source, do not reveal their data sources.
Under the hood, there are two primary OLMo 2 models – one with 7 billion parameters and another 13 billion parameter version. More parameters generally equate to more language understanding capabilities, though larger models also require more processing power to run.
According to Ai2, OLMo 2 performs on par with or better than other fully open source models of equivalent scale. They claim the 7 billion model even outperforms the popular 8 billion parameter LLama on some English language benchmarks.
OLMo 2 is designed to handle all the usual tasks – answering questions, writing essays or code, summarizing documents, and so on. The training data came from websites, academic papers, Q&A forums, and math sources.
In total, the OLMo 2 models used a massive 5 trillion token dataset during training. That’s about 3.75 billion words.
Ai2 is positioning OLMo 2 as the new gold standard for open source language AI that can be scrutinized and replicated by anyone. The company says it hopes that releasing all the pieces will “provide the open-source community with resources for new and innovative approaches.”
The full OLMo 2 models and training resources are available to download now under a permissive Apache 2.0 license.
