Available on request
Gemma API:
five sizes, from phone to server.
Google DeepMind released Gemma 4 in five sizes under Apache 2.0, with text and image input on every model and audio on the smaller ones. Tell us which size fits your workload, and we’ll talk through adding Gemma to the Xinference Model API.
Gemma 4 sizes
Pick a Gemma 4 size
| Model | Parameters | Context | Input |
|---|---|---|---|
| Gemma 4 E2B | 2.3B effective | 128K tokens | Text, image, audio |
| Gemma 4 E4B | 4.5B effective | 128K tokens | Text, image, audio |
| Gemma 4 12B Unified | 11.95B | 256K tokens | Text, image, audio |
| Gemma 4 26B A4B | 25.2B total, 3.8B active | 256K tokens | Text, image |
| Gemma 4 31B | 30.7B, dense | 256K tokens | Text, image |
Source: Google DeepMind’s Gemma 4 model card, checked 23 October 2026. All figures are the model maker’s. Gemma, Google and the Gemma logo are trademarks of Google LLC, shown only to identify the model. Their use does not imply endorsement.
Reading the names
What “E”, “A” and “Unified” mean
E for effective
E2B and E4B use per-layer embeddings, so the parameters doing the work are far fewer than the stored total. That is what makes them light enough for phones and laptops.
A for active
26B A4B is a Mixture-of-Experts model: about 4B of its parameters run for each token, so it answers almost as fast as a 4B model.
Unified for no separate encoders
The 12B Unified model feeds image patches and audio straight into the language model instead of through separate encoders, which Google says cuts multimodal latency.
Gemma API questions
Is Gemma 4 free for commercial use?
Gemma 4 is released under Apache 2.0, which permits commercial use. Earlier Gemma generations used Google’s own Gemma terms instead, so check the licence of the exact version you choose.
Which Gemma model handles audio?
E2B, E4B and 12B Unified accept audio, for speech recognition and speech-to-translated-text. The 26B A4B and 31B models take text and images only.
How many languages does Gemma 4 support?
Google lists multilingual support for over 140 languages across the Gemma 4 family.
Can I get a Gemma API on Xinference?
Yes, on request. The Model API’s standard models are Kimi K3 and DeepSeek V4.1 Flash, with public per-token pricing. For Gemma, tell us the size and expected volume.
Last updated