Ai2’s new small AI model outperforms similarly-sized models from Google, Meta

“The week is for small AI models.

Thursday, AI2, Non -Profit Research Institute AI, released Olmo 2 1B, a model of 1 billionaire parameters that AI2 claims to be wins models of similar size from Google, Meta and Alibaba on several reference values. The parameters, which are sometimes called weights, are the inner components of the model that direct its behavior.

Olmo 2 1B is available under the Apache 2.0 license on AI Dev platform hug. Unlike most models, Olmo 2 1b can be repeated from scratch; AI2 provided the code and data sets (Olmo-Mix-184,, Dolmino-Mix-1124) used it for development.

Small models may not be as capable as their Behemoth colleagues, but what matters, they do not require that the hardware be neglected. This makes them much more affordable developers and hobbies who are opposed to the limit of machines for lower and consumers.

The last few days have come up with a raft of small models, from Microsoft’s Phi 4 family family to Qwen’s 2.5 Oman 3b. Most of them – and Olmo 2 1B – can be easily launched on a modern laptop or even a mobile device.

AI2 says Olmo 2 1b is dressed at a set of 4 trillion of chips from publicly available, generated and handmade sources. Tokens are raw bits of data models that swallow and generate – 1 million tokens are equivalent to about 750,000 words.

On the reference measurement of arithmetic explanation, GSM8K, Olmo 2 1b reaches better than Google Gemma 3 1b, Meta’s Llam 3.2 1b and Alibaba Qwen 2.5 1,5B. Olmo 2 1b also eclipse of the performance of these three models on Trutilifulq, a test for assessing the factual accuracy.

AI2 warns that Olmo 2 1b, however, carries risks. Like all AI models, it can produce “problematic results”, including harmful and “sensitive” content, says the organization, as well as in fact incorrect statements. For these reasons, AI2 recommends against OLMO 2 1B deployment in commercial settings.