The Institute of Foundation Models (IFM), the research arm of the UAE's MBZUAI, released K2 Horizon on September 3, 2026: a family of six models spanning 0.9B to 375B parameters, shipped not just as open weights but with training code, training data, and full methodology under the Apache 2.0 license. IFM calls it the largest fully open model release to date.

Try it this week

Every size is downloadable now from Hugging Face, Ollama, and Unsloth, with day-zero support in vLLM and SGLang. The six sizes are 0.9B, 3.7B, 7B, 32B, a 36B-A4B mixture-of-experts, and the flagship 375B-A23B. Pick the tier that matches your hardware: the 0.9B model is built for watches, glasses, and other constrained devices, the 3.7B and 7B models run on phones and laptops, the 32B handles everyday heavy use on a workstation, and the flagship 375B-A23B is a sparse mixture-of-experts model that activates 23B parameters per token and carries a 512K-token context window. One command through Ollama or the Hugging Face CLI gets you a local model you can then fine-tune with the released recipes, no gated download or waitlist in the way.

Why it matters

Most open model launches ship weights and a license, nothing more. K2 Horizon publishes the intermediate checkpoints, data recipes, training configs, and logs alongside the final weights, which makes the models reproducible and auditable rather than just runnable. For builders who need to fine-tune on their own data or ship commercially, Apache 2.0 removes the usage restrictions that gate many open-weight rivals such as Tencent's Hy4.

Key details

IFM reports state-of-the-art results at the 0.9B, 3.7B, and 7B scales, with strong reasoning, math, coding, and agentic performance across the range. API access is offered through inference partners including Cerebras, Nebius, AWS, and Compass for teams that would rather call an endpoint than host the 375B flagship themselves.

What to do next

Start with the 7B model for a local coding or agent prototype, then scale up to the 375B endpoint through an inference partner if the quality justifies the cost. If you are doing research or need a compliance trail, download the training data and checkpoints while they are fresh, since fully open releases at this scale remain rare.