Ai2 Unveils the Tiniest OLMo 2 Model, Built for Local and Efficient AI Use

Welcome to "The AI Economy," a weekly newsletter by Ken Yeung on how AI is influencing business, work, society, and technology. Subscribe now to stay ahead with expert insights and curated updates—delivered straight to your inbox.

Nonprofit AI lab Ai2 has released a new variation of its open-source language model. Previously, OLMo 2 was available in three sizes—7B, 13B, and 32B—but now there’s a 1B version, the smallest member of the model family. The company claims it outperforms Google’s Gemma 3 1B and Meta’s Llama 3.2 1B.

OLMo 2 came onto the scene in December, with Ai2 described as the “most advanced fully open language model” on the market at the time. The addition of OLMo 2 1B gives the organization a scalable model that can be used for complex applications and now for local and lightweight programs. The company shares that this 1B model should “enable rapid iteration for researchers, more local development, and a more complete picture of how our recipe scales.”

The new variant is trained on 4T tokens of high-quality data, pulling from OLMo-mix-1124 and Dolmino-mix-1124. It’s a similar process for its OLMo 2 siblings. The company then post-trained it using Ai2’s Tulu3 recipe with distillation from strong models, on-policy preference tuning, and reinforcement learning with verifiable rewards (RLVR).

OLMo 2 1B isn’t the only new release. While 1B is the base version that is the foundation for post-training, Ai2 published five additional specialized variants. The first is an Instruct version that’s optimized for chat applications. It’s followed by a supervised fine-tuned model (SFT) for chatbots, copilots, and agents; one optimized through direct preference optimization (DPO), a GGUF format for efficient on-device inference, and a version fine-tuned with reinforcement learning from human feedback (RLHF).

If you’re wondering why a small version of OLMo 2 is being released now, Nathan Lambert, a senior machine learning researcher at Ai2, explains: “We didn’t know our 1B base model was actually good enough.” He went on to state, “We thought we had to keep pushing modeling decisions for a 1B model…or other things that are suitable for small models…Small model development can be handled much differently than bigger models.”

With Ai2 now having four OLMo 2 model types, developers should be able to use it in practically every possible use case and device.

You can download all the OLMo 2 1B artifacts on Hugging Face.