What changed
MBZUAI’s Institute of Foundation Models launched K2 Horizon on 3 September 2026: six models from 0.9 billion to 375 billion parameters sharing one architecture 1. The announcement states that every model is released fully open, with weights, code, training data and methodology, and calls it the largest fully open release in AI history 1. CellCog records Apache 2.0 for weights and code, and day-zero support in vLLM, SGLang and Ollama 2.
Openness, model by model
The claim is checkable, and somebody checked it. CellCog read the published model cards and found the fleet-level headline does not hold model by model 2:
| model | what its card says today 2 |
|---|---|
| 3.7B and 7B | data, recipe and code ship now |
| 32B | a Stage 1 checkpoint, the final one still to come |
| 36B-A4B and 375B-A23B | intermediate checkpoints, data and training code will be released |
That reading covers five of the six models 2. I read the gradient as the opposite of a broken promise: the cards publish the schedule for what is missing, so the gap is declared rather than found. Moor Insights names the artifact that matters, a per-model development tree of intermediate checkpoints and post-training branches showing how a base model becomes the reasoning and agentic variants 3. That tree is what the two largest models still owe you.
Impact on your team
Decide which end of the fleet you are buying. If you want a model you can rerun, the 3.7B and 7B are the only two whose data, recipe and code are on the table today 2: a real on-device tier, and small. If you want the 375B-A23B or the 36B-A4B, you are taking Apache 2.0 weights 2 plus a published intention, which beats most open-weight drops and is still not reproducibility. Keep “fully open” out of a procurement document for those two until the checkpoints land.
The deadline is not IFM’s, it is yours: pick a date to re-read the 375B-A23B and 32B cards, because nobody notifies you when a Stage 1 checkpoint becomes final 2. Day-zero serving support 2 makes the evaluation an afternoon: run the 7B against whatever open model you serve today.