Gemma 4 26B A4B Instruct
Google DeepMind · United States
Based on published licence terms, Gemma 4 26B A4B ships under pure Apache 2.0 with no prohibited-use carve-outs — a departure from prior Gemma generations. The sparse-MoE architecture (25.2B total / 3.8B active) puts it in an ambiguous zone for EU AI Act GPAI systemic-risk classification, and US origin plus image-input support add transparency obligations that deployers should document.
Caractéristiques de la licence
- Parameters
- 25.2B total / 3.8B active
- Architecture
- Sparse MoE — 128 experts + 1 shared, 8 routed per token, hybrid sliding-window + global attention, 30 layers
- Modality
- Text + image in, text out (no audio)
- Context length
- 256K tokens
- Data cutoff
- January 2025
- Released
- 2026-04-02
Risques connus
- Systemic-risk classification ambiguity: 25.2B total vs 3.8B active parameters sits in a zone where the EU AI Office's FLOP-threshold interpretation for sparse MoE is unsettled; deployers should assume systemic-risk-tier obligations until Google's GPAI notification is confirmed.
- Image-input Art. 50 transparency: when used to analyse images of persons, GDPR Art. 9 biometric rules and EU AI Act deep-fake/synthetic-content labelling fall on the deployer — Google provides no built-in content provenance marking.
- Training-data summary is domain-level only (web, code, maths, images, audio, 140+ languages) with no dataset enumeration or opt-out mechanism.
Sources
Revu par Ali Madjaji · Dernière revue le 2026-04-17· Revu il y a 0 jourSignaler une correction