Google DeepMind releases Gemma 4, with a 31B dense model under Apache 2.0
Gemma 4 comes in five sizes from E2B to 31B, with multimodal input, 256K context on the larger models and an Apache 2.0 licence.
Google DeepMind published Gemma 4, a family of open models. The card for the 31B instruction-tuned model lists five sizes: E2B, E4B, 12B, 26B A4B and 31B. The family handles text and image input, with audio on the E2B, E4B and 12B models, and generates text. Context is up to 256K tokens on the medium models and 128K on the small ones, with support for over 140 languages.
The card says every model is designed as a configurable reasoner and that both dense and mixture-of-experts variants are offered. The smaller sizes target laptops and phones.
The model page shows an Apache 2.0 licence and links to a Gemma 4 licence page on Google's site. The weights are ungated. A technical report and launch blog are linked from the card.
This item points to the 31B card; the date used is its Hugging Face creation date, and sibling repositories were created at different times, so treat the date as approximate.
The quiet part is the licence. Earlier Gemma terms were stricter.
Sources
Editor's first pass (AI-assisted)
- Real 4.5: Ungated Apache 2.0 weights; sizes span phone to server.
- Useful 4: Reasoning modes, coding, 140-plus languages and image input across the family.
- Notable 4.5: Over 4,100 likes on the 31B card; a major vendor's open family.
- Novel 3: Dense and MoE variants in a familiar recipe; the licence move is the change.
Comments