EU AI Act — Model Cards

Version 1.0 Effective: June 5, 2026 EU AI Act GPAIS Transparency

TL;DR — The Short Version

Skylina is a General-Purpose AI System (GPAIS) provider under the EU AI Act (Regulation (EU) 2024/1689). Each AI model we use is documented below with its risk classification, training data summary, known limitations, and failure modes. All Skylina models are classified as limited risk — no high-risk systems under Annex III.

📋 EU AI Act Compliance Summary

ModelFormatProviderRisk ClassificationEU AI Act Basis
FLUX (image generation)ImageBlackforest Labs via SapiomLimited RiskArticle 53(1)(b) GPAIS
Sora 2 (video generation)VideoOpenAI via SapiomLimited RiskArticle 53(1)(b) GPAIS
MusicGen (audio generation)MusicMeta AI via SapiomLimited RiskArticle 53(1)(b) GPAIS
Claude (code generation)CodeAnthropicLimited RiskArticle 53(1)(b) GPAIS

No Skylina models fall under Annex III high-risk AI systems. All are subject to GPAIS transparency obligations (Articles 53–56).

Sora 2

Video Generation 🟢 Limited Risk — GPAIS
Provider
OpenAI (via Sapiom)
Use in Skylina
Text-to-video, image-to-video generation
Training Data
Internet-scale video and image data through late 2024 (per OpenAI disclosures)
EU AI Act Article
Article 53(1)(b) — GPAIS
Output Format
MP4 video, up to 1080p
Watermarking
C2PA-compliant metadata embedded
CapabilityKnown Limitation / Failure Mode
Realistic motion generationMay produce physically impossible motion; physics simulation is approximate
Object persistenceObjects may change appearance or disappear in longer clips
Text renderingSignificantly less reliable than image models for readable text in video
Fine detail preservationSmall details may blur or distort across frames
Complex scene physicsMulti-object interactions, collisions, and reflections may appear unnatural
Consistent character animationCharacter appearance may drift subtly over multi-second clips
⚠️ Deepfake risk: Video content may be visually convincing. Users must clearly label AI-generated video as AI-generated when sharing publicly under EU AI Act Article 50. C2PA metadata is embedded but may not be visible to viewers.

OpenAI Sora transparency: openai.com/sora

FLUX

Image Generation 🟢 Limited Risk — GPAIS
Model Variant
FLUX.1 [dev] + FLUX.1 [schnell] + FLUX.1 [pro]
Provider
Blackforest Labs (via Sapiom)
Use in Skylina
Text-to-image, style transfer, reference-image generation
Training Data
Licensed and publicly available image-text pairs; synthetic data; filtered for safety per Blackforest Labs policy
EU AI Act Article
Article 53(1)(b) — GPAIS
Output Format
PNG, JPEG, WebP
CapabilityKnown Limitation / Failure Mode
Text rendering in imagesHighly unreliable — significant misspellings, wrong characters, extra letters; do not use for legible text generation
Celebrity / person likenessMay produce recognizable figures from training data; check outputs for unintended likeness
Complex spatial layoutsMay confuse spatial relationships (e.g., left/right, inside/outside)
Fine text/UI elementsUI mockups, code screenshots, diagrams may have artifacts or garbled content
Style consistencySpecific brand styles or character consistency across multiple images requires precise prompting
Copyright notice: Generated images may inadvertently reproduce elements from training data. You are responsible for reviewing outputs for unintended similarity to copyrighted works before commercial use.

FLUX model documentation: blackforestlabs.ai/flux

MusicGen

Music Generation 🟢 Limited Risk — GPAIS
Provider
Meta AI FAIR (via Sapiom)
Use in Skylina
Original music, background scores, soundtracks from text prompts
Training Data
Licensed music datasets; internal Meta proprietary music collection; approximately 400K hours of licensed music (per Meta disclosures)
EU AI Act Article
Article 53(1)(b) — GPAIS
Output Format
MP3, WAV
Output Duration
Up to 4 minutes per generation
CapabilityKnown Limitation / Failure Mode
Genre-specific generationHighly specialized genres (e.g., classical counterpoint, advanced jazz) may not be generated accurately
Lyrics/singing generationNot designed for lyrical singing — primarily generates instrumental tracks; voice simulation is limited
Long-form compositionLonger pieces (>2 minutes) may lose thematic coherence; use shorter clips for best results
Commercial licensingGenerated music is an AI interpretation and may not be suitable for synchronization licensing (film/TV/advertising) without additional clearance — user bears responsibility for licensing compliance
Melody accuracyMay generate melodies that closely resemble known copyrighted works — review outputs before commercial release
Music industry notice: AI-generated music may be recognized by automated Content ID systems used by music platforms. Users are responsible for understanding platform-specific AI music policies before distribution.

MusicGen research: ai.meta.com/blog/musicgen

Claude (Code Generation Mode)

Code Generation 🟢 Limited Risk — GPAIS
Provider
Anthropic PBC
Use in Skylina
Code generation, app scaffolding, bug fixes, architecture suggestions
Training Data
Internet-scale text data through early 2025; Claude 3.7 Sonnet cutoff; Anthropic's Constitutional AI training for safety alignment
EU AI Act Article
Article 53(1)(b) — GPAIS
Output Format
Source code (all major languages), architecture docs, test files
Model Version
Claude 3.7 Sonnet (current); older versions retired within 90 days of new releases
CapabilityKnown Limitation / Failure Mode
Multi-language code generationMay generate outdated API patterns or deprecated library versions — always check against current docs
Code explanation and reviewCannot execute, test, or verify correctness — outputs are guidance, not verified code
Security vulnerability detectionMay miss subtle injection flaws, race conditions, or authentication bypass vulnerabilities
Architecture suggestionsRecommendations may not match your specific infrastructure, deployment constraints, or existing codebase
Recent library/version knowledgeKnowledge cutoff means newest libraries and frameworks may be described incorrectly or missing
Long code generationVery long files may have inconsistencies or subtle bugs — prefer modular, short functions
⚠️ Production code responsibility: Generated code may contain security vulnerabilities, bugs, or outdated library versions. Test all code in a staging environment before deploying to production. Skylina is not responsible for bugs in generated code.

Anthropic Claude transparency: anthropic.com/about/policies | docs.anthropic.com

EU AI Act Obligations

Skylina is a provider of General-Purpose AI Models (GPAIS) under EU AI Act (Regulation (EU) 2024/1689, in force August 2026). We comply with GPAIS obligations:

High-risk classification: Skylina does not fall under EU AI Act Annex III (high-risk AI systems) as it does not target the fundamental rights applications listed therein (biometrics, critical infrastructure, education, employment, law enforcement, etc.).

Energy disclosure: Skylina relies on external model providers (OpenAI, Blackforest Labs, Meta AI, Anthropic) for computation. Energy consumption data is reported by those providers; Skylina does not operate dedicated GPU infrastructure.

Model Change Policy

When we switch to a materially different AI model (different provider, significantly different capabilities, or known training data changes), we update this page within 30 days and document the change in our Transparency Report. Minor model updates (bug fixes, efficiency improvements from the same provider) are noted in release notes.

Contact

Questions about our AI models or EU AI Act compliance: ai-transparency@skylinaai.com

© 2026 Skylina, built with Polsia.