Gemma models you can run locally
Google's Gemma models bring research from Gemini into compact open models for text, vision and on-device use.
Available models
Ordered from the lightest model to the largest.
Gemma 3 1B
1B · Dense
Google's tiny Gemma for on-device
32K ctx
Gemma 3 4B
4B · Dense
Multimodal Gemma with 128K context
128K ctx
Gemma 4 E2B IT
5B · Dense
Gemma 4 efficient instruct model (official)
256K ctx
Gemma 4 E4B IT
8B · Dense
Gemma 4 balanced instruct model (official)
256K ctx
Gemma 3 12B
12B · Dense
Multimodal Gemma with 128K context
128K ctx
Gemma 4 12B IT
12B · Dense
Gemma 4 mid-size instruct — multimodal any-to-any
256K ctx
DiffusionGemma 26B-A4B IT
26B · 4B active · MoE
Discrete diffusion MoE — 1100+ tok/s on H100, multimodal (text/image/video)
256K ctx
Gemma 4 26B-A4B IT
27B · 4B active · MoE
Gemma 4 MoE instruct model (official)
256K ctx
Gemma 4 31B IT
33B · Dense
Gemma 4 flagship instruct model (official)
256K ctx
Gemma 4 31B
33B · Dense
Gemma 4 flagship base model (official)
256K ctx