The Institute of Foundation Models released K2 Horizon on September 3, 2026, a connected fleet of six foundation models: 375B-A23B, 36B-A4B, 32B, 7B, 3.7B, and 0.9B. A matching press note from Abu Dhabi said the drop includes weights, code, training data, and methodologies, and called it the largest fully open-source model launch in AI history. Models and code are under Apache 2.0. Datasets use their own licenses, such as ODC-BY, with recipes disclosed when the files cannot be redistributed.

IFM said the 0.9B, 3.7B, and 7B models set new state of the art in their size classes on reasoning, mathematics, coding, and agentic tests. The 0.9B model scored above 48 on AIME 2026 and is meant for watches and glasses. The 3.7B and 7B models target phones. A dense 32B and a sparse 36B-A4B sit in the local-hosting band. The 375B-A23B flagship activates about 23 billion parameters per token for enterprise work.

Open weights, plus the training tree

The lab’s claim is not only a leaderboard. For every size it is publishing intermediate checkpoints, training data or data-construction recipes, architecture details, mixture compositions, training code, configurations, logs, evaluations, and final weights, including the agentic post-training stage. The six models share core architecture, vocabulary, methods, interfaces, and tooling, with a smaller vocabulary on the 0.9B. IFM said that makes routing and size-hopping practical.

Two technical bets sit under the scores. Mixture-of-Value-Attention, or MoVA, is a sparse attention design used on the 36B-A4B, which activates about 4 billion parameters per token. A “diffusion distillation” method generates blocks of tokens in parallel and, IFM said, speeds the models by roughly 3 times without a quality drop.

Eric Xing, founder of IFM and president of Mohamed bin Zayed University of Artificial Intelligence, framed the release as open science rather than another open-weight dump.

“Open source is much more than open weights. Science works when others can see the data, follow the method, reproduce the result, and improve on it.” Eric Xing, founder, Institute of Foundation Models

Hector Liu, director of IFM’s Silicon Valley lab, said developers can prototype on the smallest model and scale to the flagship without changing deployment workflows. The models are available on Hugging Face and through vLLM and SGLang. An API is live with inference partners Compass, Cerebras, and Nebius. IFM, launched by MBZUAI in May 2025, now has labs in Abu Dhabi, Silicon Valley, and Paris. Its other public lines include Jais, an Arabic LLM, and PAN, a world model for embodied reasoning.

Decoded Take

K2 Horizon is an argument aimed at every lab that ships a PDF and a single checkpoint. If the training trees are actually complete, researchers can study how tool use and planning appear instead of reverse-engineering a blob. The risk is the usual open-science tax: the 375B-A23B still needs serious cluster time, and “available on Hugging Face” is not the same as reproducible on a university budget. Watch whether independent teams recreate the 0.9B and 7B scores from the published recipes, whether MoVA shows up outside IFM, and whether the fleet stays coherent once the community starts forking only the sizes that fit on a workstation.