K2 Horizon just shipped as six new fully open models — developers aren’t fully convinced
The New Stack Adrian Bridgwater ● Covered by 2 sources
IFM just released six K2 Horizon AI models and says they’re fully open. Developers like the idea, but some of the training details still aren’t there yet.
Based on reporting by The New Stack, Adrian Bridgwater — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
The Institute of Foundation Models, based in Abu Dhabi, has put out K2 Horizon: six AI foundation models ranging from 0.9 billion parameters to 375 billion. IFM is calling it the largest fully open-source fleet of AI models it has released so far, and it’s using “fully open” to mean a lot more than downloadable weights.
The company says it wants to publish the whole training trail: code, evaluation code, data where it can be shared, construction recipes where it can’t, plus configurations, logs and checkpoints from pretraining through agentic post-training. That matters because IFM’s pitch is inspectability. Developers should be able to see how the models were built, reproduce parts of the process, and adapt the result.
But the launch is not as complete as the slogan suggests. All six models came with downloadable weights, yet the model cards for the 0.9B, 32B and 375B models said some data, code or checkpoints would show up later. The 32B release was also only a Stage 1 checkpoint, not the finished model. The 3.7B and 7B models were the cleanest launch: they shipped with weights, recipes, training code and data.
IFM founder Eric Xing says open source means much more than open weights, and he frames K2 Horizon as “open science.” Hector Liu, who directs IFM’s Silicon Valley lab, has also argued that the group is doing “360 degree” open source. The broader package is substantial: shared architecture, vocabulary, training methodology, interfaces, deployment tooling, and a dynamic routing system meant to send tasks to the most cost-effective model.
Still, the loudest criticism is about the bits that make reproduction real. Nitish Garg of CellCog says the compute details are missing — no accelerator count, no hours, no cost — and that the fine-grained logs may or may not fill the gap later. Others point to the same old problem with many open-weight releases: you can download the result, but the training recipe often remains a black box. K2 Horizon improves on that, but it hasn’t yet escaped the usual pattern.
My take — AI-written commentary, not fact-checked reporting
This is a real step toward openness, not the usual open-weights cosplay. But the missing training details are exactly where the magic trick lives, and AI companies know it. “Fully open” is a nice phrase until the logs, prompts, and compute story are still in the back room with the good snacks.
Read more about this at: The New Stack
