A frontier lab tied to Abu Dhabi’s MBZUAI is claiming a record for open-weight AI. The Institute of Foundation Models, launched by the university in May 2025, has put out a K2 Horizon family of six models that it calls the largest fully open-source release in the field’s history.
The range stretches from a 375B-parameter mixture-of-experts flagship with 23B active parameters down through a 36B-A4B tier to dense 32B, 7B, 3.7B and 0.9B checkpoints. Licensing is Apache 2.0 across the board, and Hugging Face hosts every size with FP8 and GGUF conversions ready to run.
Deployment support arrived the same day on vLLM, SGLang and Ollama, covering NVIDIA, AMD and Cerebras hardware, while hosted access is sold through Compass, Cerebras and Nebius. All six members share a core architecture, vocabulary and training approach, so a team can prototype on the 3.7B model and move to the flagship later without rebuilding its serving stack.
The data recipe is as public as the models. Each size trained on about 20 trillion tokens, of which roughly 17 percent were problem-solving traces that show their reasoning, and close to half was machine-written. Instruction data was folded into the middle of pre-training instead of being reserved for a final polishing pass.
IFM did not stop at weights. It also published the pre-training corpus, intermediate checkpoints, training code, configs and detailed logs, giving outsiders a full view of how the fleet was built and the raw material to extend any of its members.
