Allen Institute for AI (AI2) With its latest release the AI model is expected to take advantage of the growing demand for customized models and enterprises seeking greater transparency.
AI2 has made the latest addition to its OLMO family of large language models available to organizations, continuing to focus on openness and customization.
Olmo 3 has a longer context window, more logical 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 complete transparency and control over training data and checkpointing.
Ai2 will release three versions of Olmo 3:
-
Olmo 3-Think Both 7B and 32B are considered major logic models for advanced research
-
Olmo is also 3-base in both parameters, which is ideal for programming, comprehension, mathematics, and long-context reasoning. Ai2 said this version is “ideal for continuous pre-training or fine-tuning.”
-
Instructions in Olmo 3-7B optimized for instruction following, multi-turn dialogue, and tool use
The company said that Olmo 3-Think is “the first fully open 32B thinking model that generates explicit logic-chain-style content.” OLMO-3 Think also has a long context window of 65,000 tokens, perfect for long-running agentic projects or reasoning over long documents.
Noah Smith, AI2’s senior director of NLP research, told VentureBeat in an interview that many of its customers, from regulated enterprises to research institutions, want to use models that give them assurance about what happened in training.
“The releases from our friends in the tech world are great and super exciting, but there are a lot of people for whom data privacy controls over what goes in the models, how the models are trained and how the models can be used are front of mind,” Smith said.
Developers can access models on Hugging Face and AI2 Playground.
Transparency and customization
Smith said that with models like Olmo 3, the company believes any organization using its models will have control over it and adapt it to the way that works best for them.
“We don’t believe in one-size-fits-all solutions,” Smith said. It is a known thing in the world of machine learning that if you try to build one model that solves all problems, it is not actually the best model for any one problem. There’s no formal proof of it, but it’s something that old-timers like me have seen.”
He said models with the ability to specialize are “not as attractive as getting high scores on math tests” but offer more flexibility for enterprises.
Olmo 3 allows enterprises to essentially retrain models by adding them to the data mix they are learning from. The idea is that businesses can bring in their proprietary sources to guide the model in answering specific company questions. To help enterprises during this process, Ai2 added checkpoints from each key training phase.
The demand for model customization has increased as enterprises that cannot build their own LLMs seek to create company-specific or industry-focused models. like startup Archie to pass started climbing Enterprise-focused, customizable smaller models.
Models like Olmo 3 also give enterprises more confidence in the technology, Smith said. Because Olmo provides 3 training data, Smith said enterprises can be confident that the model hasn’t assumed anything it shouldn’t have.
Ai2 has always claimed to be committed to greater transparency, even launching a tool Olmotres in April Which can track the output of a model directly on the original training data. The company releases open-source models and posts its code to repositories like GitHub for anyone to use.
Competitors like Google and OpenAI have Developers faced criticism On moves that hide raw logic tokens and opt to summarize the logic, claiming that they no longer resort to “debugging blind” without transparency.
Ai2 pre-trained Olmo 3 on a six-trillion-token OpenAI dataset, Dolma 3. The dataset includes web data, scientific literature, and code. Smith said he optimized Olmo 3 for code, compared to the focus on mathematics for Olmo 2.
How does it stack up
AI2 claims that the OLMO 3 family of models actually represents a significant leap forward for open-source models, at least for open-source LLMs developed outside China. The base Olmo 3 model was trained “with approximately 2.5 times higher compute efficiency, measured by GPU-hours per token”, meaning it consumed less energy during pre-training and also cost less.
The company said the Olmo 3 model outperformed other Open models such as Stanford’s Marin, LLM360’s K2, and Apertus, although AI2 did not provide data for benchmark testing.
“Of note, Olmo 3-Thinking (32B) is the strongest fully open reasoning model, narrowing the gap to the best open-weight models of similar scale, such as the Qween 3-32B-Thinking series of models in our suite of reasoning benchmarks, all trained on 6x fewer tokens,” AI2 said in a press release.
The company said that Olmo 3-Instruct outperformed Quen 2.5, Gemma 3, and Llama 3.1.