Abu Dhabi's IFM ships six models with the training data attached

On 3 September the Institute of Foundation Models, launched by Mohamed bin Zayed University of Artificial Intelligence in May 2025, released K2 Horizon: six foundation models from 0.9 billion to 375 billion parameters, all under Apache 2.0. What distinguishes the release is not the size range but what ships alongside the weights. Every model comes with its training data, data recipes, training code, model configurations, intermediate checkpoints, training logs and evaluation results. That is a deliberate rebuke to the industry's prevailing usage of the word 'open'. Founder Eric Xing framed it plainly: open source is much more than open weights, and science works when others can see the data, follow the method, reproduce the result and improve on it. The smallest model is built for watches and glasses; the 3.7B and 7B models run on phones; the 32B and 36B-A4B models target laptops and on-premise servers; the 375B-A23B flagship is aimed at enterprise reasoning. All six share an architecture, vocabulary, training methodology and deployment tooling, so a developer can prototype small and scale up without changing workflow. Two technical contributions are claimed: a diffusion-distillation technique that generates token blocks in parallel for roughly a threefold speedup, and a mixture-of-value-attention architecture that improves reasoning without extra computation. The institutional reading matters more than the benchmarks. A lab funded by a Muslim-majority state, operating from Abu Dhabi, Silicon Valley and Paris, has chosen full reproducibility as its competitive position rather than a closed frontier model. Reproducibility is what lets a university in Lahore or Kano audit a model instead of merely licensing one. The models are on Hugging Face and served through vLLM and SGLang.

This is a QeRN summary by Ahmed Qerni. Read the original at Institute of Foundation Models (MBZUAI): https://ifm.ai/k2/press-release/.