IFM releases K2-Horizon open models from 7B to 375B MoE

IFM, the Institute of Foundation Models, released a new family of open models under the K2-Horizon name, according to a post on r/LocalLLaMA. The lineup spans a 7B dense model, a 32B dense model, a 36B A4B mixture-of-experts model, and a 375B A23B MoE model. The 32B is described as a stage-one checkpoint, not a finalized release. Smaller 0.9B and 3.7B models appeared a few days earlier. The models are available on Hugging Face.

The release extends IFM's K2 line into a full range of sizes, giving local and self-hosted users options from sub-1B to hundreds of billions of parameters. The 375B A23B active-parameter MoE design points at a model that routes through roughly 23B parameters per token, a configuration aimed at balancing capability with inference cost. Because the 32B is an early checkpoint, final weights for that size are still pending. Details on training data, benchmarks, and licensing were not included in the announcement.