Ai2 Launches Olmo 2 1B, Outperforming Major Competitors in AI Benchmarking

Ai2 has released the second version of its artificial intelligence model Olmo 2 1B. This small 1 billion parameter model beats comparable models from much larger tech company models on arithmetic reasoning tasks. Trained on a window of 4 trillion tokens, the new model has illustrated impressive new leaps in what AI can do.

Olmo 2 1B’s training dataset was drawn from an enormous variety of sources. It combines publicly available information, AI-generated data, and manually produced content. This rich dataset has allowed the model to get the state of the art performance on benchmark tests explicitly meant to test arithmetic reasoning. Olmo 2 1B stands tall above the competition—Google’s Gemma 3 1B, Meta’s Llama 3.2 1B, and Alibaba’s Qwen 2.5 1.5B—on the GSM8K benchmark. This amazing performance showcases its cutting edge analytical proficiency.

Olmo 2 1B Blackbird is really good at math. That’s not all. On the TruthfulQA benchmark, it beats all of its competitors, proving once again that it’s a real contender in the AI world. Ai2’s release of such a model is an excellent accomplishment for the company. Beyond that, it’s a huge historic achievement for the entire AI community.

Olmo 2 1B is publicly available under a permissive Apache 2.0 license on the AI development platform Hugging Face. This availability enables any developer or researcher to easily play with the model. They are able to run it on modern laptops and even mobile devices! Ai2 takes the guesswork out of replicating the model. It releases the code and datasets used for its development, including Olmo-mix-1124 and Dolmino-mix-1124. Ai2 has taken the extraordinary step of committing to transparency by uploading intermediate checkpoints from each 1000 steps throughout the training process.

Ai2 has already recognized the remarkable abilities of Olmo 2 1B. They caution against the dangers of deploying it in commercial environments. The company cautions against using the model as-is, urging careful consideration before deploying the model in any real-world uses.