Open source just shipped the training data, not only the weights. ifm's k2 horizon, out september…
open source just shipped the training data, not only the weights. ifm's k2 horizon, out september 3, releases six models up to 375b under apache 2.0 with code, recipes and methodology attached.
the reproducible model is the new frontier claim.
Context
The Institute of Foundation Models' blog and press release of September 3, 2026 announce K2 Horizon, six models of 375B-A23B, 36B-A4B, 32B, 7B, 3.7B and 0.9B parameters, with up to 375B being total parameters and about 23B active. The blog says the models and code are released under the Apache 2.0 license and that datasets are released under their applicable licenses, such as ODC-BY. It says the release offers training data or detailed data-construction recipes, and describes about 20 trillion tokens of pretraining per model, with intermediate checkpoints, mixtures, training logs and post-training artifacts. The 375B model card shows an apache-2.0 license and links pretrain and midtrain data repositories.
The apache license covers the weights and code and the data carries separate licenses, and the release says data or recipes, so shipped the training data is partial on the vendor's own wording. The model card, a later window, uses the future tense that the data, training code and intermediate checkpoints will be released, and the linked dataset pages and any checkpoint or code repository were not opened, so artifact availability today is unverified and no reproduction was run. Fully open and largest fully open-source launch in history are vendor superlatives without a comparative basis, and the benchmarks are vendor-reported charts. Open source just shipped the training data is the author's framing.
Related work
- K2 Horizon shipped six models from 0.9b to 375b ↗An earlier note on the same release.
- The institute of foundation models shipped k2 horizon today ↗Another earlier note on the release day.
- Mbzuai released k2 horizon ↗A later note on the same release under the MBZUAI name.
Watch next
- The dataset repositories and the training code release.
Sources
Provenance
The note above is reproduced unedited from the original post, first published on Threads on 22 September 2026 at 10:58 IST. Sources are the papers and datasets the note draws on.
View the original post ↗Embed this note
More notes
The air is now being asked to keep its own ledger
the air is now being asked to keep its own ledger: ecmwf’s aifs compo becomes the first ai model to forecast atmospheric composition globally every three hours, cleanair simulates 365 days of pm2.5 over china in ten seconds, and a unified framework maps six pollutants at one kilometer across the whole country. the air now files its own composition report.
read the note →The current is now being asked to draw its own map
the current is now being asked to draw its own map: china’s langya 2.0 predicts six ocean phenomena including internal waves and mesoscale eddies, a deep net called wenhai resolves eddies globally with air sea flux formulas built in, and scripps infers surface currents from the way temperature patterns deform in satellite images. the ocean now files its own circulation report.
read the note →The soil is now being asked to report its own carbon
the soil is now being asked to report its own carbon: a nix color sensor paired with generative data augmentation predicts soil organic carbon without a lab, random forest drives 74 percent of soil health mapping studies, and sentinel 2 tracks five year carbon change across france and italy from 922 samples. the dirt now files its own carbon account.
read the note →