Google · Open-weight
Checked as of 2026-10-23
Gemma
Google's open models: small, efficient, and Apache-2.0 from Gemma 4 on.
At a glance
- Licence
- Open source Apache-2.0 (Gemma 3: Gemma Terms of Use)
Gemma 4 is Apache-2.0 and ungated. Gemma 3 and earlier stay under the Gemma Terms of Use, with a prohibited-use policy and click-through gating.
- Access
- HFBedrockSageMaker JumpStart
- Hugging Face
- huggingface.co/google
Variants 6
| Model | Params | Context | Hardware tier | Gated | Inputs | Status |
|---|---|---|---|---|---|---|
Gemma 4 31B
google/gemma-4-31B-it
Dense. On Bedrock and JumpStart. Size from HF safetensors metadata. |
31.3B | 262.144K | 80GB GPU ml.p4de.24xlarge / ml.p5.48xlarge | No | text, image | current |
Gemma 4 26B-A4B
google/gemma-4-26B-A4B-it
Mixture of experts. On Bedrock and JumpStart. |
25.8B3.8B active | 262.144K | 80GB GPU ml.p4de.24xlarge / ml.p5.48xlarge | No | text, image | current |
Gemma 4 12B (Unified)
google/gemma-4-12B-it
Added 2026-05-23. Video is read as frames. |
11.9B | 262.144K | 40GB GPU ml.g5.12xlarge / ml.p4d.24xlarge | No | text, image, audio, video | current |
Gemma 4 E4B
google/gemma-4-E4B-it
4.5B 'effective' parameters, built for on-device. On JumpStart. |
8B | 131.072K | 40GB GPU ml.g5.12xlarge / ml.p4d.24xlarge | No | text, image, audio | current |
Gemma 4 E2B
google/gemma-4-E2B-it
2.3B 'effective' parameters, built for on-device. On Bedrock. |
5.1B | 131.072K | 16GB GPU ml.g4dn.xlarge / ml.g5.xlarge | No | text, image, audio | current |
Gemma 3 27B
google/gemma-3-27b-it
Gemma Terms of Use. |
27.4B | 131.072K | 80GB GPU ml.p4de.24xlarge / ml.p5.48xlarge | Yes | text, image | legacy |
"—" means the figure could not be verified from a primary source on 2026-10-23 and has been left blank rather than guessed. Hardware tier is from total parameters (all experts for mixture-of-experts models), since every weight must sit in memory. SageMaker rates on the open-weight page.
Strengths
- Apache-2.0 with no click-through terms from Gemma 4 on.
- E2B, E4B and 12B take audio as well as text and images.
- 26B-A4B activates under 4B parameters per token.
- Google publishes official quantised checkpoints.
Watch out for
- Context is 128K on E2B and E4B, but 256K on the larger sizes.
- E2B and E4B totals include per-layer embeddings that can be kept off the GPU, so they run on smaller hardware than their tier suggests.
- Gemma 3 is still gated under the older terms.