Everything but the number — K2 Horizon published the corpus, the checkpoints and the logs, and left out the FLOPs

The Institute of Foundation Models released K2 Horizon on 3 September 2026 - six models from 0.9B to 375B parameters under Apache 2.0, each pre-trained on roughly 20 trillion tokens, published with training code, configurations, intermediate checkpoints, fine-grained logs and evaluation results - and stated no training compute figure anywhere: no floating point operation count, no accelerator count, no hours. The obvious reading, that a maximally open release minimises a downstream team's obligations, inverts on the one clause that decides the question. Article 53(2) disapplies Article 53(1)(a) for free and open-source models with public parameters and architecture, 53(1)(a) is the Annex XI technical documentation, and Annex XI Section 1 point 2 is where the computational resources used to train the model, expressed as a number of floating point operations, would be recorded - so the omission is lawful precisely because the release is open. The Commission's guidelines of 18 July 2025 then make a downstream modifier the provider of a general-purpose AI model where the modification's training compute exceeds a third of the original model's, and where the original figure is unknown the fallback is a third of 10^23 FLOP, roughly 3.3 x 10^22 - an absolute bar rather than a share of a 375B, 20-trillion-token run. The practitioner moves are to read the model card for a FLOP figure before committing to a base, to log one's own fine-tuning compute in floating point operations from the first run, to separate per model and with a date what is published from what is promised - the 375B card still says the data and training code "will be released" - and to remember that 53(2) never lifted the copyright policy at 53(1)(c) or the public training-content summary at 53(1)(d).

Post

The most open frontier model release yet left out one number. It is the number that decides whether your fine-tune makes you a provider.

On 3 September the Institute of Foundation Models — MBZUAI's lab in Abu Dhabi — released K2 Horizon: six models from 0.9B to 375B parameters, Apache 2.0, roughly 20 trillion pre-training tokens each. Weights, training code, configs, intermediate checkpoints, fine-grained logs, evaluation results.

No compute figure. Not in the announcement, the blog, or the model cards.

Three things before you file this under "open licence, fewer obligations".

  1. The omission is lawful. AI Act Article 53(2) disapplies 53(1)(a) and (b) for models released under a free and open-source licence with public parameters and architecture. 53(1)(a) is the Annex XI technical documentation — and Annex XI Section 1 point 2 is exactly where "the computational resources used to train the model, expressed ... as a number of floating point operations" would sit. Open the model, close the file.

  2. Which hands you the fallback. The Commission's GPAI guidelines of 18 July 2025 make a downstream modifier the provider of a general-purpose AI model where the modification's training compute exceeds a third of the original model's. Where the original figure is unknown, the fallback is a third of 10^23 FLOP — roughly 3.3 × 10^22. An absolute bar, not a share.

  3. So the bar moved towards you. A third of a real 375B, 20-trillion-token run is a number no fine-tune of mine approaches. A fixed 3.3 × 10^22 is one a serious fine-tune can.

And read the dates. "Fully open" is honest as a direction: on day one the data and intermediates are public for 3.7B and 7B; the 375B card still says they "will be released".

Monday:

  1. Before you commit to an open base, check the card for a FLOP figure. That number, not the licence, sets your threshold.
  2. Log your own fine-tuning compute in FLOPs from run one. It is the only evidence you stayed under.
  3. Separate what is published from what is promised, per model, with a date.
  4. 53(2) never lifted 53(1)(c) or (d) — the copyright policy and the public training-content summary are yours either way.

Apache 2.0 tells you what you may do with the model. The number they left out tells you what you become.

Which base did you actually check?

#EUAIAct #OpenWeights #AgenticAI #AIGovernance #MachineLearning

First comment

Sources, in the order they carry the argument.

IFM's own release post — the six models, ~20T tokens each, and the artefact list: https://ifm.ai/blog/k2/

The press release, with the organisation and its sites named: https://ifm.ai/k2/press-release/

The 375B model card — Apache 2.0, 512K context, and "will be released": https://huggingface.co/IFM/K2-Horizon-375B-A23B

An independent read of the openness gap, and of the missing compute: https://cellcog.ai/blog/k2-horizon/

AI Act Article 53 and Annex XI, consolidated text: https://eur-lex.europa.eu/eli/reg/2024/1689/oj/eng

The Commission's GPAI guidelines of 18 July 2025, where the one-third criterion and the fallback sit: https://digital-strategy.ec.europa.eu/en/policies/guidelines-gpai-obligations

One point cut for length: Article 54 makes a non-Union provider appoint an authorised representative in the Union before placing a GPAI model on the market — and 53(2) does not lift that either.

Claims

Sources

Corrections

None.

AI disclosure

The video for this edition contains AI-generated media: the narration is a synthetic voice and the cover art is generated. The script, the claims and the source verification are human-authored and human-checked. Assets carry IPTC DigitalSourceType and C2PA Content Credentials.