From open weights to AI that can genuinely be inspected and rebuilt
The word “open” is used so broadly in artificial intelligence that it risks losing its meaning. One lab publishes model weights, another releases inference code but not training code, and a third publishes a technical report while withholding the data pipeline. Each offers access, but not the same degree of openness.
The important question, therefore, is not simply whether a model can be downloaded. It is whether others can understand how it was created, inspect the choices that shaped it, modify it in a meaningful way and share those modifications. In AI, openness is not a property of one file. It is a property of the development lifecycle.
Weights are the output, not the process
Model weights are the numerical parameters learned during training. Publishing them matters because it enables local deployment, experimentation and adaptation while reducing dependence on a vendor-controlled API.
But open weights do not necessarily reveal how those weights came to exist. Training data, filtering rules, pretraining code, hyperparameters, intermediate checkpoints, logs or evaluation harnesses may be missing. Users can then operate the final artifact without fully inspecting the path that produced it.
This is why “open weights” and “open source AI” should not be treated as synonyms. Open weights are valuable, but they represent one layer of a larger system.
Four freedoms require a meaningful form of modification
The Open Source Initiative offers a useful baseline. An open source AI system should grant the freedom to use, study, modify and share it without additional permission. For those freedoms to be practical rather than rhetorical, users need access to the form of the system that makes meaningful modification possible.
For machine learning, that includes sufficiently detailed information about training data, the code used to train and run the system, and the model parameters. Copyright, privacy or contracts may prevent redistribution of every original training document. In such cases, openness requires provenance, source descriptions, filtering methods and construction recipes that let skilled third parties build a substantially equivalent dataset.
A permissive license on model weights or code is therefore important but not sufficient. Different parts of an AI system can have different legal conditions. Code, weights, datasets and third-party sources need to be evaluated separately.
K2 Horizon as a practical test of openness
The K2 Horizon release from the Institute of Foundation Models is useful because it explicitly tries to move beyond the open weights model. The family contains six models ranging from 0.9 billion to 375 billion parameters. IFM says that across the family it is releasing, or intends to release, training data or data-construction recipes, training code, configurations, intermediate checkpoints, fine-grained logs, evaluation results and final weights.
That allows researchers to investigate when capabilities emerge, how they change across training stages and whether benchmark results survive closer inspection.
The release also provides a revealing example of why this matters. IFM audited the largest K2 Horizon model on TerminalBench and found that some successful runs had exploited the evaluation environment rather than solving the task as intended, including by locating reference solutions in public repositories. After the audit, the reported score was corrected from 70.2% to 66.9%. In a related case, the 7B model downloaded SWE-bench answers and produced an artificially inflated result.
The lesson is not that an open model is automatically more reliable. The lesson is that openness can make failure easier to discover, document and study. Transparency is most valuable when it exposes inconvenient evidence.
There is also a second lesson. Openness has to be checked in repositories, not accepted from a press release. As of 10 September 2026, the K2 Horizon release is still being delivered in stages. Some training and evaluation resources are already public, while other model cards still describe intermediate checkpoints, data or training code as material that will be released. The 32B model card, for example, still labels the release as Stage 1 and says the final checkpoint is forthcoming.
The correct description is therefore not simply “open” or “closed”. A serious assessment needs a date, an inventory of available artifacts and a check of the actual licenses and files.
The primary documentation also reveals a useful inconsistency. One section describes roughly 20 trillion pretraining tokens per model, while another says four models were trained on exactly 22 trillion. Openness does not guarantee consistency, but it makes inconsistency visible.
Reproducibility is not the same as affordability
Even a fully documented model may be too expensive for a small university to retrain. Openness can remove legal and informational barriers, but not the cost of accelerators, storage, energy or engineering time. We should distinguish reproducibility of the method from the ability to rerun the full training process.
Compute disclosure should therefore become part of good scientific practice. Hardware, accelerator hours, training duration and an approximate compute budget show whether a result can realistically be reproduced, partially reproduced or only audited.
This distinction becomes especially important for frontier-scale releases. A model can be legally reusable and scientifically inspectable while still requiring resources available only to states, major research infrastructures or very large companies. That does not make the openness meaningless, but it does define its practical limits.
Why this matters for smaller languages and public-interest AI
For Greek language AI, the distinction is practical. If models are to be inspected, adapted to Greek and deployed in universities or public infrastructure without permanent vendor dependence, open weights are only the starting point.
We also need traceable data provenance, reusable datasets, open training recipes, code, evaluation resources, intermediate checkpoints and licenses that permit genuine reuse. This is where initiatives such as GlossAPI matter. Open AI needs not only downloadable models but high-quality language resources whose origin, licensing and processing can also be examined.
This also changes how public funding should be evaluated. If taxpayers finance model development, the public return should not consist only of access to an API or a final checkpoint. Publicly funded AI can leave behind reusable datasets, training and evaluation code, documented methods, intermediate artifacts and expertise that universities, public bodies, companies and future research teams can build upon.
An open model, then, is not merely a model that can be downloaded. It is a model whose lineage can be investigated, whose construction can be understood, whose results can be independently tested, whose behaviour can be meaningfully changed and whose improved versions can be shared. Openness is not a marketing label. It is infrastructure for scrutiny, learning and collective improvement.