MBZUAI launches K2 Horizon open-source AI model fleet

MBZUAI’s Institute of Foundation Models (IFM) has released K2 Horizon, a connected fleet of six open-source AI models. The models range in size from 0.9 billion to 375 billion parameters, designed to serve various deployment environments including wearable devices, smartphones, laptops, and enterprise servers.

The release is characterized by its high level of transparency. Unlike standard model releases that often only provide weights, IFM is opening the entire training lifecycle. This includes intermediate checkpoints, training code, configurations, fine-grained logs, evaluation results, and detailed data-construction recipes or datasets. The models and code are released under the Apache 2.0 license.

Key technical features include a new Mixture-of-Value-Attention (MoVA) mechanism in the 36B-A4B model, which optimizes capability per active parameter. The fleet includes specialized models such as the 0.9B version for constrained devices and the 375B-A23B model for large-scale reasoning and agentic workloads. A dynamic model-routing system is also included to allow developers to transition between model sizes seamlessly.