The Allen Institute for Artificial Intelligence (Ai2) hopes to capitalize on increased demand for custom models and enterprises seeking greater transparency around AI models in its latest release.
Ai2 has made the latest addition to the Olmo family of gigantic language models available to organizations, continuing its focus on openness and customization.
Olmo 3 has a longer context window, more reasoning traces, and is better at coding than its previous iteration. This latest version, like other Olmo releases, is open source under the Apache 2.0 license. Enterprises will have full transparency and control over training data and milestones.
Ai2 will release three versions of Olmo 3:
-
Olmo 3-Think in both 7B and 32B are considered flagship reasoning models in advanced studies
-
Olmo 3- Base also in both parameters, which is ideal for programming, comprehension, mathematics and long-context reasoning. Ai2 stated that this version is “ideal for further initial training or fine-tuning
-
Olmo 3-Instruct in 7B optimized for instruction execution, multi-turn dialogue and tool use
The company stated that Olmo 3-Think is “the first-ever fully open 32B thinking model that generates explicit content in a chain of reasoning style.” Olmo-3 Think also has a long context window of 65,000 tokens, ideal for long-running agent projects or inferring from longer documents.
Noah Smith, senior director of NLP research at Ai2, told VentureBeat in an interview that many customers, from regulated enterprises to research institutions, want to use models that give them confidence in what is included in the training.
“The publications from our friends in the tech world are very cool and extremely exciting, but there are a lot of people who have control over data privacy in terms of what goes into the model, how the models are trained, and other restrictions on how the model can be used as a front-of-mind,” Smith said.
Developers can access models on Hugging Face and Ai2 Playground.
Transparency and personalization
Smith said models like the Olmo 3 that the company says every organization using their models needs to have control over and shape in a way that’s best for them.
“We don’t believe in one-size-fits-all solutions,” Smith said. It is a known fact in the world of machine learning that if you try to build a model that solves all problems, it turns out that it is not the best model for any single problem. There is no formal evidence for this, but it is something that elderly timers like me have observed.
He added that models with the ability to specialize “may not be as lightning-fast as getting high scores on math exams,” but they offer companies more flexibility.
Olmo 3 allows enterprises to essentially train a model by adding to the set of data it is learning from. The idea is that companies can leverage their proprietary sources, which will guide the model in responding to specific company queries. To lend a hand enterprises with this process, AI2 has added checkpoints for each major phase of training.
The demand for model customization has increased as companies that cannot build their own LLMs want to create company- or industry-specific models. Startups like it Arcee To have he started offering Customizable compact models designed for enterprises.
Models like the Olmo 3, Smith said, also give businesses greater confidence in the technology. Because Olmo 3 provides training data, Smith said enterprises can be sure the model hasn’t ingested anything it shouldn’t have.
Ai2 has always said it strives for greater transparency, even launching a tool called OlmoTrace in April that can trace model output directly back to the original training data. The company open-sources its models and publishes its code on repositories such as GitHub for anyone to utilize.
Competitors like Google and OpenAI have was met with criticism from developers over movements that hid raw reasoning tokens and chose to summarize the reasoning, claiming that they are now resorting to “blind debugging” without transparency.
AI2 pre-trained Olmo 3 on the OpenAI Dolma 3 dataset containing six trillion tokens. The dataset includes web data, scientific literature, and code. Smith said they optimized Olmo 3 for code compared to the math focus of Olmo 2.
How it works out
Ai2 says the Olmo 3 family is a significant step forward from truly open source models, at least for open source LLM solutions developed outside of China. The base Olmo 3 model trained “with approximately 2.5 times the compute performance as measured by GPU hours per token”, meaning it used less power during pre-training and was cheaper.
The company said the Olmo 3 models outperformed other open models such as Stanford’s Marin, LLM360’s K2 and Apertus, although the Ai2 did not provide benchmarking numbers.
“Notably, Olmo 3-Think (32B) is the strongest fully open-weighted inference model, narrowing the gap against the best open-weight models of similar scale such as the Qwen 3-32B-Thinking series of models in our inference benchmark suite, all while training on 6x fewer tokens,” Ai2 said in a press note.
The company added that Olmo 3-Instruct performs better than Qwen 2.5, Gemma 3 and Llama 3.1.