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

‘The week for small AI models, it seems.

Thursday, AI2, Non -Profit AI Research Institute, Have been released Olmo 2 1B, 1-Bilion-Parameter Model, AI2, claimed that the same size models were beaten in several benchmarks from Google, Meta and Alibaba. Parameters, sometimes known as weight, are the internal elements of a model that guides its behavior.

Olmo 2 1b AI Dave Platform is available under a permit Apache 2.0 license in the face of hugs. Unlike most models, Olmo 2 1B can be replicated from scratch; AI 2 has provided code and data sets (Olmo -Mix -1124, Dolmino -Mix -1124) Used to develop it.

Small models may not be able to be as enabled as their Behamat Parts, but important, bee hardware is not required to run them. This makes them more accessible by fighting for their developers and hobbies with the limits of low-end and consumer machines.

Microsoft’s PHI 4 logic has been launched from the family to the past few days, starting small models Qwen’s 2.5 Omni 3BThe Most of it – and Olmo 2 1b – can easily run on a modern laptop or even a mobile device.

AI2 says that Olomo 2 1B was universally trained in the data set of 4 trillion token from AI-exposed and manually made sources. Tokens are injected and generated by the raw bits of data models – 1 million token is equivalent to about 750,000 words.

In a benchmark of arguing argument, GSM 8K, Olmo 2 1B Google’s Jemma 3B, Meta Lama 3.2 1B and Alibaba Queen scored better than 2.5B. Olmo 2 1B also accepted the performance of these three models in Satyaphulka to evaluate truthful accuracy.

AI2 warned that Olmo 2 1B carries the risk. Like all the AI ​​models, it can produce “problematic output” with “sensitive” content, the company says, as well as true wrong statements. For this reason, AI2 recommends the Olmo 2 1B deployment in commercial settings.