“It’s a small AI model week.
On Thursday, AI2, a non-profit AI lab, released the Olmo 2 1B, a 100 million parameter model that AI2 claims will beat similarly sized models from Google, Meta and Alibaba on several benchmarks. A parameter, sometimes called weights, is the internal component of the model that derives its behavior.
OLMO 2 1B is available under the Apache 2.0 license that will hug the face of the AI DEV platform. Unlike most models, the Olmo 2 1B can be replicated from scratch. AI2 provided the code and dataset (Olmo-Mix-1124, Dolmino-Mix-1124) that was used to develop it.
Smaller models may not be as capable as the Giants’ counterparts, but the important thing is that they don’t need any enhanced hardware to run. This makes it more accessible to developers and enthusiasts who fight low-end and consumer restrictions.
Many small models have been launched over the past few days, from Microsoft’s Phi 4 Reasoning Family to Qwen’s 2.5 Omni 3b. Most of these – and the Olmo 2 1B can be easily run on modern laptops and mobile devices.
According to AI2, the Olmo 2 1B was trained with a dataset of 4 trillion tokens from published, AI-generated and manually created sources. A token is a raw bit of the data model that is ingested and produced. One million tokens amount to about 750,000 words.
In benchmarks that measure arithmetic inference, the GSM8K, the OLMO 2 1B outperform Google’s Gemma 3 1B, Meta’s Llama 3.2 1B, and Alibaba’s Qwen 2.5 1.5b. The Olmo 2 1b also overturns the performance of these three models in Truthfulqa, a test to assess de facto accuracy.
TechCrunch Events
Berkeley, California
|
June 5th
Book now
However, AI2 warns that its OLMO 2 1B is at risk. Like all AI models, it can generate “problematic outputs” that contain harmful and “sensitive” content. For these reasons, AI2 recommends deploying OLMO 2 1B in a commercial setting.