A model can be downloadable and still leave most of its history hidden.
On September 3, the Institute of Foundation Models, part of Mohamed bin Zayed University of Artificial Intelligence, released K2 Horizon, a family of six artificial-intelligence models ranging from approximately 0.9 billion to 375 billion parameters.
The release matters less because another set of weights appeared online than because IFM is attempting to expose substantially more of the process that produced them.
The institute says K2 Horizon will include not only final model weights, but intermediate checkpoints, training data where licensing permits, data-construction recipes where it does not, architecture information, code, configurations, logs, and evaluation results.
That is a broader definition of openness than simply allowing a finished model to be downloaded.
It is also not yet the same thing as complete reproducibility.
The finished model is only the last page
Most public discussion of open AI compresses several different ideas into one word.
Open weights mean the trained parameters can be downloaded.
Open code means portions of the software used to train, adapt, evaluate, or deploy the system are available.
Open data means researchers can inspect at least some of the material used during training.
Open data recipes can expose how datasets were selected, filtered, combined, or reconstructed when the original material cannot legally be redistributed.
Intermediate checkpoints reveal earlier states of a model during training.
Logs and configurations can show how the process was run.
None of those, individually, guarantees reproducibility.
Reproducibility requires someone outside the original laboratory to take the available materials and reproduce a meaningful part of the training process or its reported results.
That has not yet been demonstrated for K2 Horizon.
What is actually public
K2 Horizon is not merely a press release.
IFM's public Hugging Face collection contains the six principal model sizes, including 0.9B, 3.7B, 7B, 32B, 36B-A4B, and the 375B-A23B flagship.
Associated datasets and additional model resources are also visible, and IFM maintains public code repositories related to model pretraining and post-training.
The flagship model card describes the 375B-A23B system as a sparse mixture-of-experts model containing approximately 375 billion parameters while activating about 23 billion per token.
Those are concrete artifacts.
But the same official model card also preserves an important launch-day limitation.
It says the final checkpoint has been released while intermediate checkpoints, data, and training code will be released.
That future tense matters.
IFM's launch materials describe K2 Horizon as fully open and speak broadly about releasing the training lifecycle. The live artifact documentation shows that at least some parts of that broader program were still forthcoming when the models launched.
The Record therefore does not treat "fully open" as a completed factual state merely because the developer uses the phrase.
What is established is narrower:
IFM released substantial model and supporting resources immediately and committed itself publicly to exposing much more of the development process.
Transparency can be promised. Reproduction must be demonstrated.
That distinction changes how this release should be judged.
A laboratory can make itself more transparent by publishing artifacts.
It cannot declare itself independently reproducible.
That second judgment belongs to outsiders.
Can researchers reconstruct the training data mixtures from the published recipes?
Can they match benchmark results using the released evaluation code and configurations?
Can they examine intermediate checkpoints and observe when particular capabilities appeared?
Can they recreate meaningful portions of pretraining or post-training without relying on undocumented institutional knowledge?
Those questions now have a better chance of being answered because K2 Horizon provides more material to inspect.
But the answers do not exist simply because the material was promised.
Six models, one openness claim
IFM presents K2 Horizon as spanning six principal sizes, from models intended for comparatively constrained deployment environments to a 375B-parameter flagship.
It also reports that its 0.9B, 3.7B, and 7B models achieve state-of-the-art results in their respective size classes.
The Record has not independently reproduced those benchmark claims.
They remain IFM's reported results.
The same caution applies to IFM's description of K2 Horizon as the largest fully open-source model release in AI history.
Historical superlatives are difficult to establish cleanly because model families differ in parameter count, architecture, licensing, data disclosure, code availability, checkpoint publication, and what developers mean by "open."
The more useful fact is not whether K2 Horizon wins a semantic contest over the word open.
It is that IFM has made the scope of its openness claim unusually measurable.
If the promised checkpoints, data, recipes, code, logs, and evaluation materials appear, anyone will be able to compare the promise against the repository.
If they do not, that gap will also be visible.
Why the recipe matters
The finished weights of a model tell us what the model became.
They reveal much less about how it became that way.
Training data shape what patterns a system encounters.
Intermediate checkpoints can show when abilities or failures emerge.
Post-training methods alter how a system responds, reasons, refuses, or follows instructions.
Evaluation choices determine which abilities are measured and which are ignored.
Code and configurations reveal decisions that may disappear behind a final benchmark score.
For researchers trying to understand why artificial systems behave as they do, those materials are not administrative debris.
They are part of the scientific object.
K2 Horizon's significance therefore lies upstream of the final model.
The release asks whether a large model family can be treated not only as a product to use, but as a process outsiders are allowed to inspect.
Why this enters the Record
The history of artificial intelligence will eventually contain countless model releases.
Most will be remembered only if they change what systems can do, how they are used, or how they are built.
K2 Horizon may matter because it challenges a common compromise in contemporary AI: publishing a model's finished weights while keeping much of the development process private.
IFM has placed a broader promise on the table.
That promise is now testable.
Future researchers can ask whether the intermediate checkpoints arrived, whether the training data and recipes were sufficiently complete, whether the code matched the released models, whether reported benchmarks could be reproduced, and whether other laboratories followed.
The Record does not yet know the answers.
That uncertainty is part of the event.
What changed today
Six new models were released.
That alone would not make this a lasting historical marker.
What changed is that a model developer publicly attached those systems to a broader claim about how much of their construction should be visible.
Final weights and supporting resources are already public.
Some promised pieces were still described as forthcoming at launch.
Independent reproduction has not yet been demonstrated.
So the historical record is not that the entire recipe arrived with the model.
It is that the recipe itself became part of the release promise.
Now the promise can be audited.