Weights are not enough, they opened the data too: a six model fully open fleet from Abu Dhabi, the largest at 375 billion parameters

The Abu Dhabi based Institute of Foundation Models (IFM) has released K2 Horizon, a fleet of six models ranging from 0.9 billion to 375 billion parameters. According to the announcement dated 3 September, all of them are fully open: not just the weights, but the code, the training data and the methodology as well.

That distinction is the point. The "open weights" the industry has been debating in recent weeks means only the weight files are shared, while the training data and recipe stay closed. IFM founder and MBZUAI president Eric Xing puts the difference this way: "Open source is much more than open weights. Science works when others can see the data, follow the method, reproduce the result, and improve on it."

From a watch to an enterprise server

The six models split up as follows:

• 0.9B: for highly constrained environments such as watches and glasses. IFM claims it is the industry's best performing model of its size in math, reasoning and tool use.
• 3.7B: a fine tuning friendly size for phones, with the claim of the best reasoning under 4B.
• 7B: claimed as the best performing model under 10B, strong on software engineering and deep research, small enough to run on a phone.
• 32B dense: for local hosting on laptops and on premise servers.
• 36B with 4B active: a new Mixture of Value Attention architecture said to outperform many larger models while activating only 4B parameters.
• 375B with 23B active: the flagship, built for enterprise reasoning and agentic work.

All six share a core architecture, vocabulary, training methodology and interfaces, with a smaller vocabulary for the 0.9B. A developer can prototype on the small model and scale to the large one without changing deployment workflows.

Two technical claims

The announcement highlights two. The first is "diffusion distillation", a technique that generates blocks of tokens in parallel and is said to speed the model up roughly 3x without degrading response quality. The second is "mixture of value attention", an architecture said to improve reasoning without adding computation.

Hector Liu, director of IFM's Silicon Valley lab, describes the fleet approach: "Rather than release a single model, we're releasing an entire fleet at once. Developers can prototype on the smallest model, scale to the flagship, and verify every claim we make along the way."

Where to get it

The models are available immediately through Hugging Face, as well as vLLM and SGLang. On the API side, Compass, Cerebras and Nebius are listed as inference partners. Models and code are released under Apache 2.0.

Who released it

IFM was launched in May 2025 by Mohamed bin Zayed University of Artificial Intelligence (MBZUAI) and now works out of three offices, in Abu Dhabi, Silicon Valley and Paris. Alongside the K2 series, its portfolio includes Jais, an Arabic large language model, and PAN, a world model for embodied reasoning. MBZUAI, founded in 2019, has nearly a thousand students from more than 50 countries.

Why it matters

For a company that has to keep its data in house, the question is not only how good a model is but how auditable it is. When the training data and recipe are open, an organization can see what the model was trained on, show documentation against compliance questions, and reproduce the model with its own data. With a model whose weights are open but whose data is closed, none of that is possible.

The timing is notable in its own right: the same week, NVIDIA's acquisition of the open model hub Hugging Face was confirmed.