Abu Dhabi's IFM Releases Six Fully Open AI Models, Data and Training Code Included

Most "open" AI releases these days are open in name only. You get the weights, maybe a technical report with some benchmark charts, and not much else. IFM, the Abu Dhabi-based research institute, just did something different: it released six models, K2 Horizon, along with the training data, the code, the methodologies, and the intermediate checkpoints that document how each model got built.

That last part matters more than it might sound. Checkpoints are snapshots of a model at various points during training. Having them means outside researchers can actually retrace how a model evolved, not just poke at the finished product. IFM founder Eric Xing told Reuters the release lets researchers "retrace the models' development process and reproduce results." That's a much higher bar than most releases clear.

Why "Open Weight" Isn't the Same as Open Source

There's an important distinction buried in this story, and it's one worth spelling out for anyone who's been loosely using "open source" to describe any AI model you can download.

Open-weight releases, the approach favored by several Chinese developers, give you the trained parameters and let you run or fine-tune the model. What you don't get is the training data, the preprocessing pipeline, or the exact recipe used to get from raw data to finished weights. You can use the model. You can't really study how it came to be, and you definitely can't reproduce it from scratch.

Then there's the fully closed approach, which is where OpenAI and Anthropic sit. No downloadable weights, no data disclosure, no methodology papers beyond marketing-friendly summaries. You interact with these models through an API and that's the extent of your visibility.

IFM's release sits in a third category entirely, one that's genuinely rare at this scale:

• Model weights, so you can run and fine-tune the systems directly
• Training data, so you can see exactly what the models learned from
• Code and methodologies, so you can reproduce the training pipeline
• Intermediate checkpoints, so you can trace the training trajectory step by step

That combination is what separates "you can download this" from "you can understand and rebuild this." Reproducibility has been a persistent headache in AI research. Plenty of papers claim results that outside labs can't replicate because critical details, especially around data curation, get left out. A release like this closes a lot of that gap by design.

The model family itself spans an unusually wide range. On one end there's a model small enough to run on smartwatches and other constrained devices, the kind of hardware where every megabyte and every milliwatt counts. On the other end sits a 375-billion-parameter model built for enterprise deployment, the kind of scale where you're talking serious inference infrastructure and serious training budgets. Covering that entire spectrum in a single open release, from wearables to enterprise-grade systems, is a statement in itself about what IFM thinks "open" should mean across the board, not just at one convenient point on the scale.

Xing framed the release as a deliberate reference point. "Our goal with this release is to establish a reference point for what a truly open model release can look like," he told Reuters. He also said the timing and framing were meant to speak directly to policymakers and regulators who've been wrestling with how to think about AI transparency. His argument: openness and competitive performance aren't mutually exclusive, and IFM wanted receipts to prove it.

That argument lands at a pointed moment. Regulatory bodies in the US, EU, and elsewhere have spent the last couple of years debating disclosure requirements for frontier models, and the leading labs have mostly argued that competitive and safety concerns justify keeping training data and methods under wraps. A release like K2 Horizon, if the models hold up on benchmarks, undercuts that argument directly. It's hard to claim openness kills your edge when someone just published everything and is still in the conversation.

There's also a geopolitical layer here that's worth not glossing over. The launch is explicitly part of the UAE's broader push to position itself as a global AI hub, competing for relevance against the US and China in a field where compute, talent, and now apparently openness itself are all bargaining chips. Abu Dhabi has been investing heavily in AI infrastructure and research institutes over the past few years, and a release like this functions as both a technical contribution and a soft-power move: look, we can build frontier-adjacent systems and we can be transparent about it.

Reuters didn't report specific benchmark numbers for K2 Horizon in this release, so it's too early to say how the models stack up against comparable open-weight or closed competitors on standard evals. That's the natural next question anyone in the research community will be asking. Full transparency about training data and process is a meaningful contribution regardless of benchmark performance, but the "competitive performance" half of Xing's claim will need independent verification once researchers get their hands on the checkpoints.

Key Takeaways

The headline distinction here isn't just "open weights versus closed," it's a third tier that's been mostly theoretical until now: full training transparency including data and intermediate checkpoints. For researchers tired of irreproducible results, that's the real news.

A few things worth watching:

• Whether independent benchmarks confirm the models are competitive with similarly sized closed or open-weight systems
• How other labs respond, since a fully transparent release raises the bar for what "open" is expected to mean
• Whether regulators reference this release in ongoing transparency and disclosure debates
• How the 375B enterprise model performs against comparable systems from OpenAI, Anthropic, or Chinese open-weight labs

If K2 Horizon's numbers hold up under scrutiny, this could end up being cited less as a one-off release and more as the moment "open source AI" got a real definition again.