#UAE #models - The Institute of Foundation Models at Mohamed bin Zayed University of Artificial Intelligence (MBZUAI) has launched K2 Horizon, a fleet of six fully open foundation models spanning 0.9 billion to 375 billion parameters. The new model suite provides complete end-to-end transparency by publishing model weights, source code, training data, and methodology. The models range from a compact 0.9B model designed for wearable edge devices like smartwatches up to a 375B parameter flagship engineered for complex enterprise reasoning.
Architecturally, K2 Horizon incorporates diffusion distillation to generate token blocks in parallel, boosting output speeds by roughly three times, alongside a Mixture of Value Attention design and dynamic model routing. Available immediately via Hugging Face and major inference providers, the release cements the UAE’s strategy of building sovereign, fully reproducible frontier AI systems.
SO WHAT? - Releasing an entire six-model ecosystem fully open source, together with model training recipes, gives developers an end-to-end inspectable stack that challenges the closed-weight API ecosystem, while providing an immediate upgrade path from edge devices to corporate servers. The difference between the open-sourcing of the K2 model suite is the level of transparency provided. The Institute of Foundation Models calls the release “a significant step forward in transparency, well beyond the “open weights” dialogue that has dominated AI industry headlines.”
KEY POINTS:
The Institute of Foundation Models (IFM) at Mohamed bin Zayed University of Artificial Intelligence (MBZUAI) has unveiled K2 Horizon, a fleet of six AI foundation models ranging from 0.9B to 375B parameters.
Unlike typical “open weights” releases, the entire K2 Horizon family is released fully open with weights, code, training data, and methodologies under the Apache 2.0 license.
The 0.9B, 3.7B, and 7B parameter models set new performance benchmarks in their size classes across mathematics, reasoning, and coding tasks.
Architectural innovations include diffusion distillation for 3x faster token generation, Mixture of Value Attention (MoVA), and dynamic model routing across the fleet.
Deployment targets range from wearable devices and smartphones (0.9B–7B) to local servers (32B dense, 36B sparse) and enterprise infrastructure (375B flagship).
Model highlights
0.9B: the industry’s best-performing model of its size in mathematics, reasoning and tool use; compact enough to run on a watch, and suited to edge deployment in sectors including energy, logistics, healthcare and public services.
3.7B: a fine-tuning-friendly size for developers and researchers, with reasoning that is the industry’s best under 4B and matches or exceeds many larger models.
7B: the industry’s best-preforming model under 10B parameters, with strong software engineering and deep research capability; small enough to run on a phone.
32B: among the most capable dense models for local hosting; built for heavy use on laptops and on-premise servers.
36B (4B active): a new Mixture of Value Attention (MoVA) architecture that outperforms many larger models while activating only 4B parameters.
375B (23B active): the flagship, built to compete with leading open-weight models on reasoning and agentic work, and enterprise-ready.
Models are accessible immediately via Hugging Face, vLLM, and SGLang, with API access hosted by inference partners including Compass, Cerebras, AWS, and Nebius.
The launch aligns with MBZUAI’s broader national mandate, which includes training 80,000 federal employees as agentic AI experts for the UAE government.
ZOOM OUT - K2 Horizon marks the culmination of an aggressive, open-source research trajectory driven by MBZUAI and the Institute of Foundation Models since its establishment in May 2025. The institute previously established its 360-open (LLM360) methodology through earlier releases, beginning with the original K2 base model, the 32B K2-Think, the 70B sovereign K2-V2 base, and the fully sovereign K2 Think V2 reasoning model in January 2026. By expanding from single reasoning checkpoints into an interconnected, multi-scale model fleet, IFM is scaling the UAE’s sovereign AI infrastructure globally, offering researchers an alternative to proprietary API lock-in through completely inspectable and reproducible foundation models.
LINKS
Read technical blog (IFM)
K2 Horizon model code (Hugging Face)
K2 Horizon model code (Github)
IFM - In May 2025, MBZUAI launched the Institute of Foundation Models (IFM), anchored by a new Silicon Valley facility in Sunnyvale, California. Operating as a tri-continental research network spanning Abu Dhabi, Paris, and Silicon Valley, the IFM serves as a primary engine for the UAE’s economic diversification and sovereign technology strategy. The institute aims to advance frontier-class open-source AI models, such as the PAN world simulation system, the K2 reasoning architecture, and the JAIS Arabic LLM family. IFM prioritises radical transparency through its LLM360 initiative, which releases full source code, datasets, and training checkpoints. By bridging California's innovation ecosystem with international academic and industrial hubs, the IFM connects global AI talent to build scalable, inspectable foundation models for deployment across enterprise, government, and scientific applications.
[Written and edited with the assistance of AI]
Source: MBZUAI
Read more about the history of K2 models:
MBZUAI officially launches K2 Think V2 with mobile apps (Middle East AI News)
MBZUAI releases K2 V2 open-source reasoning model (Middle East AI News)
K2 Think 32B rivals reasoning models 20 times its size (Middle East AI News)
UAE President backs UAE-made AI reasoning model (Middle East AI News)