G42, Cerebras Systems, and the Institute of Foundation Models at Mohamed bin Zayed University of Artificial Intelligence said they have released K2 Think V2, a 70 billion parameter reasoning system built on the new K2 V2 base model. The organizations described K2 V2 as IFM’s strongest frontier class open source foundation model and said it was purpose built to serve as the base for K2 Think.
The groups positioned the release as a milestone for the UAE’s technological sovereignty, emphasizing that K2 Think V2 is open end to end across the full model lifecycle. They said the system’s openness spans pre training data and curation, post training, reasoning alignment, and evaluation, enabling independent inspection and reproducibility while increasing transparency around how the model is built and assessed.
According to the announcement, K2 Think V2 represents a shift from a reasoning layer added on top of an existing model to a reasoning first system built directly into the foundation model. The release is also framed as expanding long context capability to support sustained multi step reasoning across larger bodies of information.
The organizations said K2 Think V2 relies on IFM curated training datasets, including the Guru dataset, and that its data pipeline is decontaminated from downstream benchmarks to support fair evaluation. They also said the release makes available major components needed for replication and extension, including intermediate checkpoints, post training recipes, and evaluation artifacts.
In benchmarking claims, the announcement said K2 Think V2 delivers leading results on complex reasoning tests including AIME2025, GPQA Diamond, HMMT, and IFBench when compared with other open source reasoning systems. The release is presented as part of a broader effort by MBZUAI and IFM to build open, high performance foundation models and reasoning systems with an emphasis on transparency, scientific credibility, and sovereignty.
