The obligations themselves aren't new — general-purpose model providers have nominally been under the Act since August 2, 2025. What changed on Sunday is that the regulator got teeth. From now the AI Office can demand technical documentation, run its own model evaluations, order corrective measures, and pull a model from the EU market. Article 50's transparency duties bite too: if your product talks to users as a chatbot or generates images, audio, or video, you have to say so and label the output, deepfakes included. Worth being precise about the scope, because the headlines blur it. The Digital Omnibus deferred the Annex III high-risk obligations — the employment, biometric, and critical-infrastructure use cases — out to 2027 and 2028. So this is a general-purpose-model event, not the full Act landing at once. If you ship or fine-tune a frontier model for the EU market, the compliance clock is now real. If you're just integrating someone else's model, your near-term homework is the transparency labeling, not the high-risk paperwork.