The Institute of Foundation Models at MBZUAI released K2 Horizon on September 3, 2026. The family includes six models ranging from 0.9 billion to 375 billion parameters, with options designed for devices, local hosting, research, and larger enterprise systems.

What makes the release notable is not only the number of models. IFM says each model includes weights, code, training resources, checkpoints, evaluation details, and methodology under the Apache 2.0 license.

The six model sizes

  • 0.9B for compact devices such as watches and glasses
  • 3.7B for lightweight experimentation and fine-tuning
  • 7B for phones and other edge devices
  • 32B for local hosting
  • 36B with 4B active parameters for efficient inference
  • 375B with 23B active parameters for larger workloads

That range gives teams a practical path from a small prototype to a more capable deployment without changing model families at every stage.

Why fully open matters

Many releases are described as open even when only the weights are available. K2 Horizon gives researchers and developers more of the material needed to inspect how the models were built, reproduce parts of the work, and adapt the system for a specialized task.

Open access does not make deployment effortless. Teams still need to review the license, test model behavior, plan for hardware, and protect any private data used for fine-tuning. The advantage is choice. Builders are not limited to a hosted API when a local or controlled deployment is a better fit.

How builders can use it

The models are available through Hugging Face and can be served with tools such as vLLM and SGLang. That makes the family relevant to developers who already work with common open-model tooling.

A small team might start with the 7B or 32B model, measure quality on a narrow workflow, and compare local operating costs with a commercial API. The largest model will not automatically be the best choice. Latency, memory, privacy, and maintenance all matter.

The SideQuest Stack take

K2 Horizon is useful because it treats open AI as a full development stack, not a download link. The strongest opportunity is model choice across devices and infrastructure. The biggest responsibility is testing each size against a real task before committing to a deployment.

Read the MBZUAI announcement and visit the K2 Horizon product page.