Weights, code and training data are published for the entire range, from 0.9 to 375 billion parameters.
The Institute of Foundation Models (IFM), attached to MBZUAI University, presented on 3 September K2 Horizon, a family of six foundation models ranging from 0.9 to 375 billion parameters. Particularity claimed: the opening is total, weight, code and training data included, allowing anyone to inspect, reproduce and adapt models.
The smallest versions (0.9, 3.7 and 7 billion parameters) target embedded uses — connected watches, glasses, smartphones — when heavier versions, including a dense 32 billion model and a scattered 36 billion model with partial activation, are designed for local hosting or on internal servers. In terms of reasoning, mathematics, code and the use of agents, IFM claims a level of performance among the best in their size category, with new records for versions 0.9, 3.7 and 7 billion parameters.
Read the full article on the original source
Go to the source →