K2 Horizon is putting a new definition of “open AI” into the spotlight. The Abu Dhabi-based Institute of Foundation Models (IFM), part of Mohamed bin Zayed University of Artificial Intelligence, has released six AI models along with their weights, code, training data, methodologies and intermediate development checkpoints.
The September 3 release is notable because researchers can examine not only the finished models but also much of the process used to create them. IFM says this makes K2 Horizon a reproducible blueprint rather than simply another downloadable AI model.
What is K2 Horizon?
K2 Horizon is a family of six foundation models ranging from 0.9 billion to 375 billion parameters. The models are designed for different computing environments, from highly constrained devices such as smartwatches and glasses to enterprise-scale deployments.
The lineup includes:
- 0.9B — designed for extremely constrained devices
- 3.7B and 7B — aimed at phones and on-device applications
- 32B — a larger dense model for local deployment
- 36B-A4B — a sparse model activating about 4 billion parameters per token
- 375B-A23B — the flagship enterprise model, with about 23 billion active parameters per token
All six models support deployment options ranging from local hardware to large-scale infrastructure.
Why is the release different?
The biggest distinction is transparency.
Many AI companies use the term “open” when they make model weights available for download. K2 Horizon goes considerably further. IFM says it is releasing training data or detailed data-construction recipes, training code, configurations, evaluation resources, methodologies and intermediate checkpoints.
That means researchers can study how capabilities develop during training rather than examining only the final model. The models and code are released under the Apache 2.0 license, while datasets follow their applicable licenses.
Why does K2 Horizon matter?
The release arrives as the AI industry debates whether openness should mean downloadable weights or complete transparency about how models are built.
IFM founder Eric Xing argues that openness and competitive performance do not have to be opposites. The institute is effectively challenging developers to disclose more of the AI-development pipeline.
There is also a strategic dimension. The UAE has been investing heavily in artificial intelligence infrastructure, research and talent as it seeks to establish Abu Dhabi as a global AI hub.
K2 Horizon gives that ambition a technological showcase: six models covering everything from edge devices to enterprise AI, with the development process opened up for inspection.
What happens next?
The significance of K2 Horizon may ultimately depend on what researchers do with the release.
If developers can reproduce results, modify the training pipeline and build new models from the published resources, it could become an important reference point for genuinely transparent AI development.
The larger question is whether other AI labs will follow the UAE’s lead—or whether “fully open” AI remains the exception rather than the new standard.
