Six models, one release
The Institute of Foundation Models, the Abu Dhabi lab founded by MBZUAI in May 2025, released K2 Horizon on 3 September: six models at 0.9B, 3.7B, 7B and 32B parameters as dense models, plus two sparse mixture-of-experts models at 36B with about 4B active parameters and 375B with about 23B active.
IFM gives each size a job rather than treating them as one model shrunk down. The smallest targets edge hardware, the mid sizes on-device agents and local hosting, and the 375B model larger reasoning and agentic workloads. The models carry a context window of 524,288 tokens according to their model cards, and IFM says each was pretrained on roughly 20 trillion tokens.
Weights and code are Apache 2.0. Datasets keep their own licences, including ODC-BY where redistribution allows.
The claim that makes it different
Plenty of labs publish weights. IFM’s pitch is that weights alone are not open source.
“Open source is much more than open weights,” said Eric Xing, IFM’s founder. “Science works when others can see the data, follow the method, reproduce the result, and improve on it.”
The release is meant to include training code, configurations, intermediate checkpoints, fine-grained logs, evaluation results and the training data itself — with construction recipes substituted where a dataset cannot be redistributed. Hector Liu, who directs IFM’s Silicon Valley lab, framed the scale of it: “Rather than release a single model, we’re releasing an entire fleet.”
That is a materially stronger commitment than the open-weight releases the field has become used to, and it is the part worth checking rather than repeating.
What actually shipped on day one
The openness is staged. Reviewing the model cards, CellCog reports that the 3.7B and 7B models arrived with the full package — weights, data and recipes, code and intermediate checkpoints — while the flagship 375B and the 36B sparse model shipped with weights only, their cards stating that checkpoints, data and training code will be released. The 32B is labelled a stage-one checkpoint with a final version pending.
So the models most researchers would want to reproduce are, for now, the ones with the least published behind them. That is not a broken promise; it is an unfulfilled one, and the difference matters only if the rest arrives.
Performance and where to get it
IFM says the three smallest models set the state of the art in their size classes. Those figures are the lab’s own; independent evaluation has not yet caught up, and no vendor benchmark should be read as a settled result.
The weights are on Hugging Face — the platform Nvidia agreed to buy for $12.93bn the same week — and run under vLLM, SGLang and Ollama. API access is offered through Compass, Cerebras, AWS and Nebius.
What to watch next
Two things settle whether K2 Horizon is what it says it is. First, whether the training data and code for the 375B and 36B models are published, and on what timeline. Second, whether independent evaluations reproduce the size-class claims. Until both land, this is a very large open-weight release with an open-science promise attached.
