Gemma 4 E4B Instruct
Google DeepMind · United States
Based on published licence terms, Gemma 4 E4B is an edge-optimised multimodal variant under pure Apache 2.0 with no prohibited-use carve-outs. Audio input (30s) and on-device deployment push GDPR biometric, AI Act emotion-recognition, and Art. 25 data-protection-by-design obligations entirely onto the integrator with no Google-side telemetry or kill-switch.
Caractéristiques de la licence
- Parameters
- ~8B raw / 4.5B effective via Per-Layer Embeddings + ~150M vision encoder
- Architecture
- Dense, edge-optimised (PLE compression, 512-token sliding window)
- Modality
- Text + image + audio (≤30s) + video-as-frames (≤60s) in, text out
- Context length
- 128K tokens
- Data cutoff
- January 2025
- Released
- 2026-04-02
Risques connus
- On-device audio capture: 30-second audio input on laptops/mobile creates GDPR lawful-basis, Art. 9 special-category, and EU AI Act Art. 5 (emotion-recognition / biometric-categorisation) exposure entirely on the integrator.
- Edge-deployment accountability: weights run offline on user hardware, so GDPR Art. 25 data-protection-by-design and AI Act Art. 16 provider/deployer demarcation fall solely on the integrator — no Google-side telemetry, audit log, or remote disablement.
- Training-data summary is domain-level only with no dataset enumeration or opt-out mechanism — thin for Art. 53(1)(d) review even though model size likely sits below the systemic-risk threshold.
Sources
Revu par Ali Madjaji · Dernière revue le 2026-04-17· Revu il y a 0 jourSignaler une correction