Photo: Tima Miroshnichenko / Pexels
The UAE-backed research lab is betting that full transparency in AI training can compete with proprietary giants
The Institute of Foundation Models at Mohamed bin Zayed University of Artificial Intelligence has released K2 Horizon, a family of six open-source AI models ranging from 0.9 billion to 375 billion parameters.
MBZUAI, based in Abu Dhabi, has been steadily building out its K2 model family over the past year.
What the K2 family actually does #
The K2 series is built around three core capabilities: general reasoning, long-context handling, and tool utilization.
The lineage here matters. K2 Think launched in September 2025 with 32 billion parameters, designed specifically for reasoning-intensive workloads. By January 2026, the team had scaled that up to K2 Think V2 at 70 billion parameters. K2-V2, released in December 2025, earned top marks on the Artificial Analysis Openness Index, a ranking that evaluates how transparent AI labs actually are about their training data, methods, and model weights.
That openness score means the training processes, model weights, and datasets are all released under permissive licenses. Researchers can inspect the full pipeline, reproduce results, and fine-tune models for their own purposes.
The UAE’s foundation model ambitions #
MBZUAI has partnerships with G42, the Abu Dhabi-based AI and cloud computing company, and Cerebras, the chip maker known for its wafer-scale processors. Those collaborations focus on optimizing inference performance.
Technology Innovation Institute, another Abu Dhabi research lab, released the Falcon series of models. MBZUAI’s K2 family represents a parallel effort with a distinct emphasis on training transparency.
Why open-source transparency keeps gaining ground #
K2’s strong performance on benchmarks related to mathematics, science, and reasoning tasks positions it as a credible alternative to proprietary options.
The permissive licensing on K2 Horizon means companies can build commercial products on top of these models without the restrictive clauses that some nominally “open” releases include.
Disclosure: This article was edited by Editorial Team. For more information on how we create and review content, see our