IFM releases K2 Horizon, six Apache-2.0 models from 0.9B to 375B-A23B with the entire training lifecycle opened, not just weights
The Institute of Foundation Models released K2 Horizon on 2026-09-03: 375B-A23B, 36B-A4B, 32B, 7B, 3.7B and 0.9B, sharing one architecture, vocabulary, training methodology and eval infrastructure. For every model they publish intermediate checkpoints, training data or the data-construction recipe, training code, configs, fine-grained logs and eval results under Apache 2.0, including the agentic post-training stages, which no other open family has exposed. IFM claims SOTA in class for the 0.9B, 3.7B and 7B tiers, with the 0.9B scoring above 48 on AIME 2026, and the 36B-A4B using a new Mixture-of-Value-Attention mechanism to punch above its active parameter count. GGUF conversions were already appearing on Hugging Face by 09-04.
↳ Follow the thread