CLAIM-2026-09-08-03

No training compute figure appears in the K2 Horizon announcement, the release blog post or the model cards: no floating point operation count, no accelerator count, no GPU-hours, no training duration.

Why it matters

Training compute is the quantity the Commission's downstream-modifier criterion is calculated from. Its absence is what moves a fine-tuning team from a proportional test to a fixed one.

Sources

• instituteof2026introduci
• huggingface2026khorizonb
• cellcog2026k