'Fully Open' Models Are Here — and Developers Aren't Convinced. What K2 Horizon Reveals About Open-Weight AI.

On the third of September the Institute of Foundation Models in Abu Dhabi released K2 Horizon, a fleet of six models running from 0.9 billion to 375 billion parameters, and called it the largest fully open-source launch in AI history. Weights, code, training data, methodologies, the lot. Within days, the developers who actually matter had a different verdict: hold on, some of this is still arriving. The gap between those two claims is the most useful thing to happen to the open-source AI debate in months, and it has real consequences for anyone choosing between frontier APIs, self-hosted models, and the platform layer in between.

What K2 Horizon actually shipped

Let me be fair to the announcement first, because on paper it is substantial. K2 Horizon is a connected fleet of six models: a 375 billion parameter flagship with 23 billion active parameters, a 36 billion with 4 billion active, then a 32B, 7B, 3.7B, and 0.9B for smaller and on-device jobs. Everything is Apache-2.0 licensed, which for a model of this size is a real signal of intent.

The pitch is that this is not another open-weight release with a smile and no receipts. IFM, which sits under Abu Dhabi's MBZUAI, says the launch includes weights, code, training data, and methodologies, enough for a researcher to reproduce and adapt the models. That is the kind of disclosure the open-source community has been demanding from Meta, Mistral, and everyone else for years. If it held up, it would reset the bar.

Where the developers pushed back

It does not fully hold up, and the pushback was fast. Within a day of the release, the criticism had settled into three specific complaints.

First, the 32B model shipped as a Stage 1 checkpoint, not a finished model. The final version and remaining artifacts are still to come. The problem is not just the waiting. It is that the Stage 1 checkpoint trails the competition on its own published benchmarks. On Terminal-Bench, a measure of how well a model handles agentic terminal tasks, the K2 32B scored 36.6 against 79.8 for Qwen3.8-27B. Shipping an incomplete model and publishing a table that shows it losing is a strange way to win trust.

Second, compute is not disclosed anywhere. No accelerator count, no training hours, no cost. For a launch whose entire selling point is transparency, that is a large hole. Critics such as CellCog's Nitish Garg were quick to point out that without compute disclosure, you cannot reproduce the model no matter how much data and code you publish. Reproducibility is the whole point of open, and it is the one thing still missing.

Third, the bigger models are marked will be released rather than released. The 375B-A23B and 36B-A4B have their final weights out, but their intermediate checkpoints, training data, and training code are promised for later. The smaller 3.7B and 7B ship with the full set: data, recipe, code, and intermediates. So the openness is uneven across the fleet, richest where the models are smallest.

What open actually means now

None of this makes K2 Horizon a bad release. Several reviewers concluded that the 7B is quietly the strongest part of the fleet, and a fully documented small model you can run on a laptop is worth more to most builders than a 375B you cannot afford to serve. But the episode does something more useful than any single model: it exposes how the word open has been stretched until it barely means anything.

Open-source used to mean one thing. You could read the code, build it yourself, and change it. With large models that definition fell apart, because a model is not really code. It is weights, plus the data and compute that produced them. Publishing the weights without the data is like publishing a compiled binary without the source. It is usable, but it is not open in any sense a developer would have recognised ten years ago.

The industry has responded by inventing a ladder of openness. Open weights at the bottom: you can run it and fine-tune it, but you have no idea how it was made. Above that, weights plus code and a model card. Above that, data recipes and training logs. At the very top, full reproducibility: weights, data, code, and enough compute disclosure to rebuild the thing from scratch. Almost nobody ships at the top. K2 Horizon tried to, and the fact that it fell short on compute and shipped one incomplete checkpoint tells you how hard that top rung is.

Why this matters to builders, not just researchers

You might be reading this thinking it is a niche argument for people who train models, and you would be half right. But there is a practical through-line for anyone building on AI, and it is about risk.

When you self-host an open-weight model, you are making a bet that you can keep it running, fix it when it breaks, and understand it when it misbehaves. That bet only pays off if the model is actually legible. A model with published data and a recipe is something you can interrogate. A model that is a black box with a permissive licence is just a cheaper API you happen to host yourself. If you do not care about the legibility, you may as well use the frontier API and save yourself the infrastructure.

I see too many teams self-hosting an open model for the warm feeling of sovereignty while getting none of the actual benefits, because the model they picked is no more inspectable than the API they left. That is the wrong trade. The reason to go open is not ideology. It is control, cost at scale, and the ability to audit what you ship. If the model does not give you those things, you have paid the operational price of open without collecting the value.

The takeaway

K2 Horizon is a real release and an incomplete one at the same time, and that combination is the whole story. The largest fully open fleet in history turned out to be mostly open, partly promised, and silent on compute, which means even the best-resourced attempts at full openness are struggling with the hardest part.

For builders the lesson is simple and I will state it plainly. Stop treating open as a binary. It is a checklist. Before you commit to any open model, ask three questions: are the weights and licence actually usable for what I plan, is the training data and recipe published so I can understand and audit it, and is compute disclosed so the thing is actually reproducible. Two out of three is not open. It is marketing, and you should price it accordingly.