On 3 September 2026 the Institute of Foundation Models (IFM), the Abu Dhabi lab launched by MBZUAI in May 2025, introduced K2 Horizon: six foundation models from 0.9 billion to 375 billion parameters, released with their weights, code, training data, methodologies and intermediate checkpoints.[1][2]
IFM's own claim is that this is "the largest fully open-source model launch in AI history" — and the release is deliberately pitched beyond the open-weights debate.[1]
Reuters reads it the same way: the move "goes beyond the 'open-weight' approach used by some Chinese developers, which make models available for download but provide limited insight into how they were built", and contrasts with OpenAI and Anthropic, "which neither release their models for download nor disclose the data and techniques behind them".[2]
End-to-end reproducibility is the stated reason: "Science works when others can see the data, follow the method, reproduce the result, and improve on it," said Eric Xing, IFM founder and MBZUAI president. "Every model in the fleet ships with its training data, recipe and evaluations."[1]
The six models share a core architecture, vocabulary, training methodology, interfaces and deployment tooling, with a smaller vocabulary for the 0.9B model.[1] IFM's dynamic model routing directs tasks to the most cost-effective model in the fleet.[1]
Two techniques are highlighted in the announcement: a "diffusion distillation" method that generates blocks of tokens in parallel and "speeds model performance by roughly 3X without degrading response quality", and a "mixture of value attention" architecture that improves reasoning without adding computation.[1]
The models and code are released under the Apache 2.0 license, available immediately through Hugging Face, vLLM and SGLang, with the API live via inference partners including Compass, Cerebras and Nebius.[1]
The launch is part of the UAE's push to establish itself as a global AI hub, and IFM now runs labs in Abu Dhabi, Silicon Valley and Paris.[1][2]
The release is explicitly framed as a policy statement: Xing said it aims "to establish a reference point for what a truly open model release can look like", and to show policymakers and regulators that "openness and competitive performance are not mutually exclusive".[2]
It is one data point in an unusually busy week of releases — Anthropic shipped Claude Fable 5.1 and Mythos 5.1 on 1 September, Google shipped Gemini 3.8 Flash on 2 September, and Meta shipped Muse Spark 1.3 the same evening.[3]
K2 Horizon is the strongest test yet of the proposition that open source means open science, not just open weights: training data and recipes in the public domain under Apache 2.0, from a state-backed lab that wants to be a reference point.[1][2]
Whether the training-data promise survives contact with actual revisability — and whether "largest fully open launch" holds up as the fleet is audited — is the follow-up worth watching.[1][2]
[1] https://ifm.ai/k2/press-release — IFM press release: Institute of Foundation Models Launches the Industry's Largest Fully Open-Source Fleet of AI Models (2026-09-03, fetched live)
[2] https://zawya.com/en/business/ai-shaping-digital-future/abu-dhabi-ai-institute-releases-fully-open-source-models-with-training-data-code-477344 — Reuters (Leo Marchandon) via Zawya: Abu Dhabi AI institute releases fully open-source models with training data, code
[3] https://www.digitalapplied.com/blog/ai-model-releases-september-2026-tracker — digitalapplied: AI Model Releases, September 2026 Tracker (surrounding releases context)