Announcement versus repository: the open-source test for K2 Horizon

On 3 September 2026 the Institute of Foundation Models (IFM) — founded in May 2025 by the Mohamed bin Zayed University of Artificial Intelligence and operating across Abu Dhabi, Silicon Valley and Paris — announced the release of the K2 Horizon family. Six models under an Apache 2.0 licence, in sizes of 0.9, 3.7, 7, 32 and 36 billion parameters with 4 billion active (the MoVA variant), plus 375 billion with 23 billion active per token. The official technical sheet gives the 375B model a native context window of 512K tokens (524,288); one trade outlet reports 131,072 for the same model. The discrepancy has not been cleared up. On independent performance, Artificial Analysis awards the 375B model a score of 38 on its own Intelligence Index, against a median of 22 measured among open-weight models of comparable size.

The real open question, though, is how complete the release actually is. In the statement distributed on PR Newswire, the institute describes the distribution as covering weights, source code, training data and methodologies for all six variants. IFM founder and MBZUAI president Eric Xing underlined that approach: “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.” Even so, a check of the organisation's repositories on Hugging Face shows the model weights alongside a single dataset devoted to evaluation sources, named “eval-360-sources”. The model card for the 375B model itself acknowledges the staggered timing, stating that “We have released the final checkpoint; intermediate checkpoints, along with the data and the training code, will be released.” The training dataset and the training code are not published at present, which leaves the promise of full reproducibility made at launch impossible to verify. IFM has not responded publicly to the discrepancy, and no timetable for publishing the training data is known.

Available for integration with vLLM and SGLang and usable through inference partners such as Compass, Cerebras, AWS and Nebius, the models do include some notable engineering. According to the statement, the 36B MoVA variant adopts a Mixture of Value Attention architecture and a diffusion distillation technique with a claimed threefold speed-up — figures that have not been independently verified. The total number of tokens used in training and the volume of compute employed remain undisclosed. The official ifm.ai site refuses automated requests (HTTP 403): the text of the institute's blog was not read directly, and the statements reported here come from the PR Newswire release and from the model card.

Claiming to have moved beyond the open-weights-only model is an ambitious step. But as long as the training data and the original recipes remain promises in the future tense, the distance between the announcement and the scientific proof is still there to be closed.

— Olya