IFM ships K2 Horizon: six open-weight models, 0.9B to 375B-A23B

The Institute of Foundation Models (IFM, MBZUAI) released K2 Horizon on 3 September 2026. The IFM blog lists six models: 375B-A23B, 36B-A4B, 32B, 7B, 3.7B, and 0.9B. Hub cards carry Apache 2.0 tags, and the Hugging Face collection is live.

From the 3 September 2026 @IFM_AI post: "Introducing K2 Horizon: a connected fleet of six foundation models ranging from 0.9 billion to 375 billion parameters."

The blog: "Today IFM is releasing K2 Horizon, a connected fleet of six models: 375B-A23B, 36B-A4B, 32B, 7B, 3.7B, and 0.9B. Across reasoning, mathematics, coding, agentic tasks, and general capabilities, K2 Horizon delivers top-tier performance in every size class, with the 0.9B, 3.7B, and 7B models setting new state of the art at their respective scales."

Key facts: - Flagship weights: K2-Horizon-375B-A23B (MoE, about 23B active). Dense and smaller siblings live in the same collection: MoVA-36B-A4B, 32B, 7B, 3.7B, 0.9B. Cards list 524,288-token context except 0.9B at 131,072. - IFM describes MoVA (Mixture-of-Value Attention) on the 36B-A4B model: about 4B active parameters per token, slightly below the dense 32B under the same training. - The 375B model card says the final checkpoint is released; intermediate checkpoints, data, and training code "will be released." The 7B card says training data, recipe, code, and eval resources are public. Dataset tags name `IFM/K2-Horizon-Pretrain-Data` and `IFM/K2-Horizon-Midtrain-Data`. - IFM says each model was pretrained on about 20 trillion tokens, and that 3.7B, 7B, 32B, and 36B-A4B share the same 22 trillion-token run.

How to read: live Hub checkpoints, Apache 2.0 tags, and the listed sizes are the hard facts. Size-class state of the art and agentic ranks are IFM-reported. Openness is uneven across sizes: 375B still points some artifacts at a later dump.

Independence: Frontier Flux is independent of IFM and MBZUAI. Facts track the 3 September X post, the IFM blog, and Hugging Face cards.