G42 has announced the release of K2 Think V2. This latest 70-billion parameter advanced reasoning system reasoning system is now built on the K2-V2 base model, IFM’s strongest fully frontier-class open-source foundation model that’s purpose built for K2 Think.
This launch marks a major milestone for the UAE’s technological sovereignty. While earlier versions of K2 Think were open and widely accessible, K2 Think V2 (70B) is the first to be open from end-to-end; from pre-training data and curation through post-training, reasoning alignment, and evaluation. The result is a reasoning system that is both more transparent and more powerful.
By upgrading the base model to K2-V2, K2 Think V2 unlocks a new level of performance, openness, and independence, reinforcing the UAE’s leadership in building frontier-grade AI systems that are globally accessible and fully sovereign.
Why K2 Think V2 is Different
K2 Think V2 marks a step change in open reasoning systems, shifting from a reasoning model layered on top of a foundation to one built directly into it. Inheriting K2-V2’s long-context capabilities and full training transparency, it now operates as a fully sovereign system end-to-end.
Every stage of K2 Think V2 (70B) is open, inspectable, and independently reproducible, ensuring both scientific credibility and national AI sovereignty. Designed as a purpose-built reasoning system, it solves complex problems step-by-step using long chains of thought across mathematics, science, coding, logic, and simulation.
This reasoning-first, fully open foundation translates directly into greater levels of performance. K2 Think V2 delivers leading results on complex reasoning benchmarks, including AIME2025, GPQA-Diamond, HMMT, and IFBench, when compared against other open-source reasoning systems.
Key Features, Made Simple
Built on K2-V2
K2 Think V2 (70B) is powered by MBZUAI IFM’s latest foundation model, designed from the outset to support reasoning, long context, and alignment, raising the ceiling for what the reasoning system can achieve.
Longer Memory, Deeper Thought
Expanded context length enables sustained, multi-step reasoning over much larger bodies of information.
Truly Independent Data Pipeline
All training relies on IFM-curated datasets, including the Guru dataset, fully decontaminated from downstream benchmarks to ensure fair and trustworthy evaluation.
360-Open Transparency
From pre-training data and intermediate checkpoints to post-training recipes and evaluations, every component is available for inspection, reuse, and extension.
