K2 Horizon: six fully open models to 375B, with the training data
Abu Dhabi's IFM put an Apache-2.0 fleet on Hugging Face — small models ship full recipes you can reproduce; the biggest ones aren't fully open yet.
Copy markdown
Six Apache-2.0 models, 0.9B to 375B
IFM's K2 Horizon fleet runs from a 0.9B edge model to a 375B (23B active) flagship, all Apache-2.0 and live on Hugging Face. Pull them and serve on vLLM or SGLang today, or route through Cerebras, Nebius, and Compass.
The rare part: the training data ships too
Unlike most 'open weights' drops, the 3.7B and 7B come with their full training data, code, and evals. You can reproduce the run or continue pretraining from a known-good recipe instead of inheriting a black box.
The asterisk: 'fully' is still arriving
Read the model cards before you plan around them. The 375B and 36B list data and code as 'will be released,' and the 32B is only a Stage-1 checkpoint — weights you can run now, but the reproducible pipeline for the big ones isn't here yet.
Build this weekend: an on-device model that's yours
The 0.9B and 3.7B target watches, glasses, and phones with best-in-class-for-size claims on math and tool use. With the recipe in hand, fine-tuning one into a private, offline app is a weekend project, not a research program.