The Institute of Foundation Models, based in Abu Dhabi, released K2 Horizon last week, a collection of six AI foundation models that the organization describes as the "largest fully open-source fleet of AI models" available to date. The models range from 0.9 billion to 375 billion parameters, and IFM pledges to publish not just downloadable weights but also training and evaluation code, training data or detailed construction recipes, configurations, logs, and intermediate checkpoints spanning the entire development process. The goal is to allow developers to see how the models were constructed, replicate their creation, and modify them for custom applications.
The six models span different use cases and environments. The 0.9 billion parameter model is built for highly constrained devices like smartwatches and smartglasses, while the 3.7 billion and 7 billion parameter versions target phones and other on-device uses. A dense 32 billion parameter model and a sparse 36 billion-A4B model offer stronger performance for local hosting and on-premises servers, and the flagship 375 billion-A23B model delivers the fleet's most powerful capabilities for demanding enterprise deployments. All six share a core architecture, vocabulary, training methodology, interfaces, and deployment tooling, with only the smallest model using a reduced vocabulary. However, not all artifacts were available at launch—while all six models shipped with downloadable weights, the model cards for the 0.9 billion, 32 billion, and 375 billion parameter versions indicated that some training data, code, or checkpoints would arrive later, and the 32 billion release was only a Stage 1 checkpoint with the final model still to come.
"Open source is much more than open weights," said Eric Xing, IFM founder and professor at the Mohamed bin Zayed University of Artificial Intelligence, in a statement. "Science works when others can see the data, follow the method, reproduce the result, and improve on it." IFM states it is opening the full training lifecycle for every K2 Horizon model, from pretraining through reasoning and agentic post-training, with each model releasing or committing to release intermediate checkpoints, training data, or detailed data-construction recipes. The checkpoints are snapshots saved throughout training, letting researchers examine how a model develops and reproduce or resume specific stages. Hector Liu, director of IFM's Silicon Valley lab, said in a recent video interview that at IFM "we are the pioneer of 360 [degree] open source or fully open source," distinguishing the approach from releases that only open-weight the final model without explaining how training or production was done.
IFM's approach positions the release as a move toward transparency that goes "well beyond" the open-weights conversation that has dominated AI industry headlines this year. The organization emphasizes that sharing the broader artifact set—including architecture details, mixture compositions, training code, configurations, fine-grained logs, evaluation results, and final weights—represents "a significant step forward in transparency." Yet questions remain about the completeness of that openness. According to Nitish Garg, founder and CEO of AI company CellCog, reasoning traces for math were rewritten into dialogues and study guides and mixed into pretraining rather than saved for post-training, and compute is not disclosed anywhere—no accelerator count, no hours, no cost. "For a release whose thesis is inspectability, that is the one obvious hole," Garg noted, adding that the fine-grained training logs, when they arrive, may fill it. Developers discussing the release have also pointed to the need for full generator prompts, seed code, and exact filtering heuristics to achieve true end-to-end reproducibility, as well as precise details of hardware topology, distributed communication configurations, and optimizer state records. The weights are available on Hugging Face with launch-day support for vLLM and SGLang, and the K2 Horizon API is accessible through IFM's inference partners, including Compass, Cerebras, and Nebius, with all models and code released under the Apache 2.0 license.
IFM's dynamic model routing technique directs tasks to the most cost-effective model and provides developers with a path from prototype to production, allowing them to prototype on the smallest model, scale to the flagship, and verify every claim IFM makes along the way. The organization's commitment to publishing synthetic datasets, methodologies, and detailed recipes aims to enable researchers to follow the full development arc and improve upon it. The staggered availability of training artifacts means the release's ultimate impact will depend on whether the forthcoming code, datasets, and logs deliver the level of reproducibility IFM promises. If they do, K2 Horizon could set a new standard for openness in foundation model development—one that forces other labs to choose between matching that transparency or explaining why they won't. The choice between shipping finished products and exposing the messy reality of model training is no longer just a philosophical debate but a competitive differentiator that enterprises and researchers will weigh when deciding which models to trust and build upon.