It seems that this is a week for a small AI model.

On Thursday, nonprofit AI research institute AI2 released the Olmo 2 1B, a 100 million parameter model that AI2 claims beat several benchmarks from Google, Meta and Alibaba. Parameters, sometimes called weights, are internal components of the model that guides its behavior.

Olmo 2 1B is available under the Apache 2.0 license on the AI ​​Dev platform embrace. Unlike most models, the Olmo 2 1B can be copied from scratch; AI2 provides the code and datasets used to develop it (Olmo-Mix-1124, Dolmino-Mix-1124).

Small models may not be as capable as the behemoth, but importantly, they don't require tough hardware to run. This makes them easier to get with the limitations of low-end and consumer machines for developers and hobbyists.

There have been a range of small models launched over the past few days, from Microsoft's PHI 4 Inference Family to Qwen's 2.5 Omni 3B. Most of these (and the Olmo 2 1B) can easily run on modern laptops or even mobile devices.

AI2 says Olmo 2 1B is trained in datasets for publicly available, AI generation and manual creation of resources. Tokens are the original bits ingested and generated by the data model - 1 million tokens equals about 750,000 words.

In benchmarks that measure arithmetic reasoning, the GSM8K, Olmo 2 1B scores are better than Google's Gemma 3 1B, Meta's Llama 3.2 1B and Alibaba's Qwen 2.5 1.5B. Olmo 2 1b also eclipses the performance of these three models, a test that evaluates factual accuracy.

AI2 warns that the Olmo 2 1B is risky. Like all AI models, it can produce "problematic output", including harmful and "sensitive" content, as well as statements that are actually inaccurate, the group says. For these reasons, AI2 recommends not deploying Olmo 2 1B in a commercial environment.