“It’s the week for small AI models, it looks.
Non -profit AI2 Research Institute on Thursday liberated OLMO 2 1B, a 1 billion parameter model that the AI2 claims to hit models of similar size from Google, Meta and Alibaba at various points of reference. The parameters, sometimes referred to as weights, are the internal components of a model that guide its behavior.
OLMO 2 1B is available with Apache 2.0 permissible license on the AI Dev platform. Unlike most models, OLMO 2 1B can be reproduced from scratch as AI2 has provided sets of code and data (OLMO-MIX-1124 and Dolmino-Mix-1124) used to develop it.
Small models may not be as capable as their counterparts, but mainly do not require bulky material to run. This makes them much more accessible to developers and hobbyists who question the limitations of lower levels of material and consumers.
There has been a series of small models that have been launched in recent days by Microsoft’s Phi 4 Reasoning Family The 2.5 omni 3b of Qwen. Most of them, including the Olmo 2 1B, can easily run on a modern laptop or even a mobile device.
AI2 says that OLMO 2 1B was trained in a database of 4 trillion brands from available to the public created by AI and manual sources. The brands are the raw pieces of data that model and create, with a million brands equivalent to about 750,000 words.
In a reference sign that measured the numerical reasoning, the GSM8K, the OLMO 2 1B scores better than Google’s Gemma 3B, Meta’s Llama 3.2 1B, and Alibaba’s Qwen 2.5 1.5B. The Olmo 2 1B also disappears the performance of these three models in Truthfulqa, a test to evaluate the actual accuracy.
TechCrunch event
Berkeley, ca
|
June 5
Book now
AI2 warned that Olmo 2 1b is dangerous. Like all AI models, it can produce “problematic outputs”, including harmful and “sensitive” content, the organization said, as well as in truly inaccurate statements. For these reasons, AI2 recommends against the development of OLMO 2 1B in trade.
