Ai2 is putting a sharper definition behind the phrase open AI. The nonprofit AI research organization founded by the late Microsoft co-founder Paul Allen has released OLMo 2, a new family of language models built around public access to the ingredients behind the system, not just access to the finished model.
The release matters because many language models are described as open, but their training data, training code, recipes, evaluations, and development checkpoints are often not all available. Ai2 is positioning OLMo 2 as a model family that can be examined, reproduced, and reused with far more visibility into how it was made.
What Ai2 released
OLMo 2 is the second family of models in Ai2's OLMo series. OLMo stands for "open language model," and the new release includes two versions: OLMo 7B and OLMo 13B.
The numbers refer to parameters. The source describes parameters as roughly corresponding to a model's problem-solving skills, with larger models generally performing better than smaller ones. In practical terms, Ai2 is offering one model with 7 billion parameters and another with 13 billion parameters.
Like many language models, OLMo 2 can handle text-based work. The source lists examples such as answering questions, summarizing documents, and writing code. That makes the family relevant not only for AI researchers, but also for developers and organizations looking at language models as general-purpose tools.
The release also places OLMo 2 in direct comparison with other open models, including Meta's Llama. Ai2 claims the new models are competitive with Meta's Llama 3.1 release, and says the smaller OLMo 2 model outperforms Llama 3.1 8B.
Why the open-source claim is central
The important point in this release is not only that OLMo 2 can be downloaded. It is that Ai2 says the model family meets the Open Source Initiative's definition of open source AI.
The Open Source Initiative finalized that definition in October. According to the source, Ai2's first OLMo models, released in February, also met the same criterion. OLMo 2 continues that approach with a broader package of development materials.
Ai2 described the release this way in a blog post: "OLMo 2 [was] developed start-to-finish with open and accessible training data, open-source training code, reproducible training recipes, transparent evaluations, intermediate checkpoints, and more."
That list is significant because it separates model access from model reproducibility. A downloadable model can be useful, but it does not necessarily tell researchers how it was created. Ai2's argument is that the open-source community needs access to the full stack in order to verify results, repeat experiments, and build on previous work.
The company also wrote: "By openly sharing our data, recipes, and findings, we hope to provide the open-source community with the resources needed to discover new and innovative approaches."
How OLMo 2 was trained
Ai2 trained the OLMo 2 models on a dataset of 5 trillion tokens. The source explains tokens as pieces of raw data, noting that 1 million tokens is equal to about 750,000 words.
The training material included several categories of content. Ai2 used websites that were "filtered for high quality," academic papers, Q&A discussion boards, and math workbooks. The math workbooks were described as "both synthetic and human generated."
Those details matter because training data is one of the biggest differences between a model that is merely available and one that can be meaningfully studied. If developers and researchers can see the data sources and recipes, they can better understand where a model's strengths and weaknesses may come from.
Ai2 says the result is a meaningful jump over its earlier work. In its own comparison, the organization wrote: "Not only do we observe a dramatic improvement in performance across all tasks compared to our earlier OLMo model but, notably, OLMo 2 7B outperforms Llama 3.1 8B."
Ai2 also described OLMo 2 as "the best fully-open language models to date." That is Ai2's claim, and the release is clearly intended to make openness part of the performance conversation rather than a separate research value.
Commercial use and developer access
The OLMo 2 models and their components can be downloaded from Ai2's website. They are released under the Apache 2.0 license, which means they can be used commercially.
For developers, that combination is important. Commercial permission makes the models more practical for projects outside academic research, while the open development materials can help teams understand what they are adopting.
The release gives users access to more than a model file. Based on Ai2's description, the package includes the pieces needed to inspect the model's development process, including training code, training recipes, evaluations, and intermediate checkpoints.
That does not automatically make OLMo 2 the right model for every use case. But it does give the AI community another option in a market where performance, licensing, and transparency are increasingly intertwined.
The safety debate around open models
The source also notes that open models remain part of a wider safety debate. Llama models were reportedly used by Chinese researchers to develop defense tools, raising questions about how openly available models might be used beyond their creators' intentions.
Ai2 engineer Dirk Groeneveld addressed that concern in February when asked whether he worried about OLMo being abused. He acknowledged the risk, saying: "Yes, it’s possible open models may be used inappropriately or for unintended purposes."
But he argued that the tradeoff still favors openness. In the same response, he said the approach "also promotes technical advancements that lead to more ethical models; is a prerequisite for verification and reproducibility, as these can only be achieved with access to the full stack; and reduces a growing concentration of power, creating more equitable access."
That framing captures the central tension around OLMo 2. The same openness that enables verification and broader participation can also create concern about misuse. Ai2's position is that access to the full development stack is necessary for a healthier AI ecosystem.
With OLMo 2, Ai2 is not just releasing another language model family. It is testing whether a model can compete with widely used open alternatives while also making its training process visible enough for others to inspect, reproduce, and build upon.