Abu Dhabi-based research institute IFM released six AI models on September 3, 2026, bundled with training data, code, methodologies, and intermediate checkpoints, in what amounts to a full-stack open-source release1.
The K2 Horizon release allows researchers to retrace the models' development process and reproduce results, IFM founder Eric Xing told Reuters. The package includes model weights alongside the underlying datasets and training procedures, going beyond the "open-weight" approach used by some Chinese developers, which make models available for download but provide limited insight into how they were built.
The distinction matters. Open-weight releases, such as those from Deep Seek and Meta, give external developers access to a finished artifact but not the recipe. By publishing data, code, and intermediate checkpoints, IFM is enabling third parties to audit training decisions, identify data-quality issues, and fork the development process at any stage.
ANALYSIS The checkpoint-level transparency is particularly notable: intermediate snapshots let researchers study how model behavior evolves during training, a capability that pure weight releases do not support.
The release positions IFM at the far end of the openness spectrum relative to both Western and Chinese peers, where even nominally "open" projects typically withhold training data or full reproducibility tooling.
Eric Xing framed the release as a challenge to an industry trend toward increasingly secretive AI development. The six models and their accompanying artifacts are intended to let outside teams not only use the outputs but interrogate and replicate the process that produced them.
The K2 Horizon family arrives as the open-source versus closed-source debate continues to shape competitive dynamics across the AI industry. IFM's decision to publish the full training stack, rather than weights alone, sets a concrete benchmark for what "fully open-source" means in practice.