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

‘A week for small AI models, appears.

Thursday, AI2, Non-Profit AI Research Institute, released OLD 2 1B, AI2 allegations in several criteria Google, Meta and Alibaba’nin a 1 billion parameter model with similar sized models. Sometimes the parameters called weights are the internal components of a model that directs the behavior.

OLD 2 1B is available under a perisive Apache 2.0 license associated with the AI ​​dev platform. Unlike most models, OLD 2 1B can be repeated from scratch; Provided AI2 code and data sets (Olmo-Mix-1124, Dolmino-Mix-1124) used to develop this.

Small models may not be as skilled as their behemotic counterparts, but important, they do not require beef hardware. It makes them more accessible to those who fight with low levels and the restrictions of consumer machines and for hobbyists.

Over the past few days, it was a raft of a small model from Microsoft Phi 4 Reasoning Family for Qwenin 2.5 Oman 3B. Most of them – and OLD 2 1B – can easily work easily on a modern laptop or even on a mobile device.

AI2 says Ilmo 2 1B, made up of 4 trillion trillion trims from the originating and hand-created sources. Tokens are raw bits of data models and arise – 1 million Token is about 750,000 words.

Measurement of arithmetic justification, GSM8K, OLD 2 1B, Google’s Gemma 3 1B, methane Llama 3.2 1B and Alibaban Qwen 2.5 1.5B. OLD 2 1B, OLD 2 1B, which is a test to assess the accuracy of the truth, also occupies the performance of these three models.

AI2 warns OLD 2 1B’s risk. Like all AI models, it can lead to “problem speeches”, including harmful and “sensitive” content, as well as inaccurate phrases. For these reasons, the AI2 recommends OLD to the placement of 2 1b in commercial parameters.