MBZUAI's IFM released six open models with full weights, data, and training logs

The Institute of Foundation Models (IFM), part of Abu Dhabi's Mohamed bin Zayed University of Artificial Intelligence, released K2 Horizon on September 3, 2026: six models ranging from 0.9 billion to 375 billion parameters, all under the Apache 2.0 license. Per multiple outlets' reporting on its announcement, IFM is billing this as the largest fully open model release in AI history. Unlike a typical open-weights release, it says it published the pretraining datasets, training code, model configurations, and evaluation results alongside the weights.

Checking the claim against the artifacts

IFM's own press materials blocked direct access, so we checked the release on Hugging Face. IFM's organization page lists the full K2 Horizon model family plus several published datasets, including TxT360-v2 for pretraining and separate math- and code-reasoning datasets. Training data is public along with the weights. The flagship 375B-A23B model's card is more measured than the marketing: at the time we checked, it describes the final weights as released, with intermediate checkpoints and training code still forthcoming. For the largest model, what's downloadable today falls short of the release's stated scope.

A separate technical claim

IFM also claims a technique it calls diffusion distillation, which pairs a frozen autoregressive model with lightweight adapters that generate blocks of tokens in parallel, delivers roughly a 3x inference speedup with no loss in output quality, per Middle East AI News's reporting on the release. That is IFM's description of its own architecture, not an independent benchmark. Read it like any other lab's self-reported efficiency number.

The "largest fully open release" superlative is also IFM's own claim, and no source we found offers a rigorous method for ranking open releases. What can be verified is that IFM published more of its research pipeline than most labs, training data and methodology included. That matters if you're deciding whether to self-host an open-weight model or pay for API access: the more of a model's construction you can check, the easier it is to reason about its cost and behavior.