Gemma 4
GoogleMULTIMODAL Open weight
- Provider
- Modality
- MULTIMODAL
- Parameters
- 31B
- Context window
- 256,000 tokens
- Weights
- Open
- Released
- 31 Mar 2026
Gemma 4 (March 2026) is Google's current open-weight family — free, self-hostable, multimodal models from edge-sized (E2B) up to 31B, with 140+ language support.
Best for
- On-prem, edge and on-device deployment with full data control
- Text + image + audio input across 140+ languages
- E2B/E4B for mobile, 12B for laptops, 31B/26B-MoE for servers
- Free fine-tuning, zero per-token cost
How it compares — and the India angle
- The only option here you can run fully offline/on-prem — decisive for Indian firms with data-residency or cost-control needs.
- 140+ languages incl. Indic make it strong for Hindi/regional on-device apps without API bills.
- Pick a hosted Gemini tier instead when you need frontier reasoning or lack GPU infrastructure.
How to access
Free open weights (self-host / on-device). Sizes ~2B (E2B) to 31B dense; 26B is MoE (~4B active).
Access Gemma 4Frequently asked questions
What is Gemma 4?
Gemma 4 is a multimodal AI model developed by Google with 31 billion parameters.
Who made Gemma 4?
Gemma 4 was created by Google.
Is Gemma 4 open source?
Yes, Gemma 4 has open weights and can be downloaded and self-hosted.
What is Gemma 4's context window?
Gemma 4 supports a context window of 256,000 tokens.