K2 Horizon Ships Six Open Models, and the Training Data With Them

While four commercial labs were shipping frontier models last week, a research institute in Abu Dhabi released six of its own and did something the big labs do not: it published the training data alongside them. The Institute of Foundation Models, part of Mohamed bin Zayed University of Artificial Intelligence, launched K2 Horizon on 3 September under the Apache 2.0 licence, which is about as permissive as software licences get.

The fleet spans 0.9 billion to 375 billion parameters. Parameters are the adjustable numbers inside a model, and more of them usually means more capability and more hardware needed to run it. There are dense models at 0.9B, 3.7B, 7B and 32B, plus two sparse mixture-of-experts models at 36B with 4B active and 375B with 23B active. Sparse means only a fraction of the model wakes up for any given question, so a 375B model can cost closer to a 23B one to run. IFM aims the 0.9B model at watches and glasses, the 3.7B and 7B at phones, the 32B and 36B at laptops and on-premise servers, and the 375B flagship at enterprise work. IFM claims the 0.9B, 3.7B and 7B models set new state of the art at their sizes. Two technical notes from the release: a “diffusion distillation” method that generates blocks of tokens in parallel for roughly a 3x speedup, and a mixture-of-value-attention architecture meant to improve reasoning without extra compute.

Why “fully open” is not the same as “open weights”. Most models that get called open, including Llama and DeepSeek, publish the weights: the finished numbers you need to run the thing. They do not publish what went into training, which means you can use the model but you cannot check it, reproduce it, or study why it behaves the way it does. IFM founder Eric Xing put the distinction directly, arguing that science works when others can see the data, follow the method and reproduce the result. Whether the models hold up against the best open-weight releases is a question for independent testing over the coming weeks, and self-reported benchmarks from any lab deserve the same scepticism. But the transparency claim is checkable in a way benchmark scores are not, which is rather the point.

What this means for you. If you have wanted to run a capable model on your own machine rather than sending your text to someone else’s server, the 7B and 32B models are the ones to look at, and they are on Hugging Face with vLLM, SGLang and Ollama support plus GGUF builds, which is the format most desktop tools want. A 7B model runs comfortably on a decent laptop. If you are studying or teaching, the published training data makes this the most inspectable serious model family available right now, which matters if your question is “why did it say that” rather than “what did it say”. For everyone else: nothing changes today, but a fully reproducible model at this scale gives researchers a tool they have been asking for since 2023.

Sources

• K2 Horizon press release (Institute of Foundation Models)
• MBZUAI’s Institute of Foundation Models launches K2 Horizon
• IFM Releases Fully Open K2 Horizon Models with Weights, Code and Training Data (AIwire)