Google DeepMind · Open-weight
Gemma 4
Sizes E2B, E4B, 26B A4B (MoE) and 31B dense
- Provider
- Google DeepMind
- Family
- Gemma
- Model type
- Open-weight
- Official release date
- Mar 31, 2026
- Release type
- New model
- Status
- Generally available
- Previous version
- Gemma 3
- Input
- text, image
- Output
- text
- Context window
- Not documented
- Max output
- Not documented
- API model id
- Not documented
- API price (input)
- Not verified
- API price (output)
- Not verified
- Price notes
- —
- Consumer apps
- —
- Developer API
- —
- Free access
- None documented
- Paid access
- —
- Availability
- Downloadable weights (Hugging Face and others)
- Licence
- open-weight (Apache 2.0)
Main improvements
- Sizes E2B, E4B, 26B A4B (MoE) and 31B dense
- MTP variants added 2026-04-16; Gemma 4 12B Unified added 2026-06-03
- Licensed under Apache 2.0 per launch coverage
Best use cases
- On-device and self-hosted AI
- Fine-tuning
Benchmarks
No benchmark results recorded. We only list scores with a source, benchmark version and date.