MBZUAI Launches K2 Horizon Fleet to Redefine 'Fully Open' AI

The Institute of Foundation Models (IFM) at MBZUAI has released K2 Horizon, a suite of six AI models designed to challenge the industry's standard for open-source transparency. The release spans a wide range of scales, from a compact 0.9B parameter model to a massive 375B parameter sparse model.

Released under the Apache 2.0 license, the K2 Horizon fleet includes models at 0.9B, 3.7B, 7B, 32B, 36B-A4B (sparse), and 375B-A23B (sparse) parameters. To ensure immediate utility for developers, the models are supported by popular inference frameworks including vLLM, SGLang, and Ollama. Unlike many contemporary releases that only provide model weights, IFM is pitching this as a "fully open" release. According to the institute, this definition requires the provision of weights, training and evaluation code, training data or recipes, configurations, logs, and intermediate checkpoints.

The Battle for 'Open Source'

This launch arrives during a period of intense debate over what constitutes "open source" in the era of generative AI. Many leading AI labs release "open weights" models, allowing users to run the models locally, but keep the underlying training data and specific methodologies secret. This gap creates a "black box" effect that prevents independent researchers from fully understanding how a model arrived at its capabilities or identifying specific biases in the training set.

IFM is attempting to differentiate itself by practicing what it calls "360-degree open source." Eric Xing, the founder of IFM, argues that open source must extend beyond weights to be scientifically valid. Xing stated that science only works when others can see the data, follow the method, reproduce the result, and improve upon it.

Industry Implications

If IFM successfully delivers the full training lifecycle for a model as large as 375B parameters, it would establish a new transparency benchmark for the entire AI sector. Such a move would shift the pressure onto other labs to disclose their training recipes and datasets, potentially accelerating the pace of academic research by allowing scientists to build directly upon verified foundations rather than guessing at the training parameters of proprietary systems.

Remaining Hurdles

Despite the ambitious goals, the path to total transparency remains difficult. Some developers have expressed skepticism regarding the release, noting that certain artifacts for the flagship models were not immediately available at launch. Furthermore, the lack of disclosed compute costs—including the number of accelerators used and total training hours—has been highlighted as a significant gap in an otherwise transparent release.

Observers will now be watching to see if IFM fills these documentation gaps and whether the broader developer community can successfully reproduce the results of the K2 Horizon fleet using the provided data and code.