The Institute of Foundation Models (IFM) released „K2 Horizon“ on 3 September 2026, a family of six AI language models – and it opens up not only the weights but also the training data, code and training logs. The lab, founded by Abu Dhabi’s MBZUAI university, calls it the industry’s largest fully open-source model fleet.
Six sizes from smartwatch to data centre
The fleet spans models of 0.9, 3.7, 7, 32, 36 and 375 billion parameters. IFM aims the smallest variants at heavily constrained devices such as smartwatches and smart glasses, as well as smartphones. The dense 32B model and the sparse 36B-A4B model suit local servers, while the flagship K2-Horizon-375B-A23B is a Mixture-of-Experts that activates just 23 billion parameters per token. It offers a 512K-token context window.
Open all the way down to the training logs
Unlike most open releases that stop at the weights, IFM additionally publishes training data, configurations, intermediate checkpoints and training logs – all under the permissive Apache 2.0 licence. The models are available on Hugging Face and run on the vLLM and SGLang inference frameworks.
The vendor’s benchmark numbers
For its flagship model, IFM cites the following self-reported scores:
- GPQA Diamond: 87.3 percent
- SWE-Bench Pro: 42.6 percent
- Terminal-Bench 2.1: 70.2 percent – revised down by the vendor to 66.9 percent after a reward-hacking audit
- Humanity’s Last Exam (no tools): 32.0 percent
The figures come from the maker and have not yet been independently confirmed. Notably, IFM revised its own Terminal-Bench score downward after flagging 24 test runs as faulty – an unusually transparent move in a field that rarely spares optimistic self-reporting.
Sources: IFM blog · Hugging Face model card · PR Newswire
