XenoSpectrum Examines AI Act Rules for Modified Models
Whether low-compute changes that remove AI safeguards bring model modifiers within the EU AI Act’s provider rules is the focus of a XenoSpectrum analysis published on November 20, 2026. The analysis examines how the compute-based criteria apply to safeguard removal and redistribution, without establishing that those activities fall outside the rules.
European Commission guidance makes provider status dependent on whether a modification significantly changes a general-purpose model’s capabilities, generality or systemic risk. Training compute exceeding one-third of the amount used for the original model is an indicator, rather than an automatic test. If the original amount cannot reasonably be established or estimated, the guidance gives fallback values: one-third of 10^25 FLOP for systemic-risk models and one-third of 10^23 FLOP for other models.
XenoSpectrum illustrates the comparison with a four-billion-parameter model. Its estimate puts safeguard removal at approximately 6.8×10^17 FLOP, assuming an RTX 3090 operates for 40 minutes at its theoretical sparse peak performance of 284 TFLOPS. That is approximately one fifty-thousandth of a hypothetical benchmark of 3.3×10^22 FLOP, calculated as one-third of assumed original training compute of 10^23 FLOP. The analysis identifies this as a theoretical calculation, not measured compute, and leaves unresolved whether the operations qualify as training compute under the Act.
The provider obligations described in Regulation (EU) 2024/1689 include model documentation, subject to a conditional open-source exemption. According to the Commission’s FAQ, general-purpose model providers ordinarily must keep technical documentation and supply information to downstream AI-system providers. Some documentation requirements are exempted for qualifying open-source models whose parameters are public and whose free licence permits access, use, modification and distribution without monetisation. The FAQ says copyright-policy and training-data-summary requirements still apply, while systemic-risk models are excluded from that exemption.
Redistribution is also covered by a separate 10a Labs analysis described by XenoSpectrum. Of 11,635 uncensored-model repositories identified on Hugging Face between January 2024 and March 2026, it classified 8,164, approximately 70%, as redistributed copies, often involving quantization or comparable compression. That analysis used GPT-5 for classification and omitted merged models and copies hosted elsewhere.
A different study discussed by XenoSpectrum counted 22,096 uncensored open-source language models using public metadata. Its dataset card identifies five covered hubs: ModelScope, Kaggle, Hugging Face, John Snow Labs and OpenCSG. According to lead author Zilong Lin, as reported by XenoSpectrum, approximately 13% of the surveyed models had a moderation warning. That figure concerns the presence of labels in the collection, rather than testing how each model behaves.
* This article is based on publicly available information at the time of writing.