---
title: "Gemma 4 26B A4B API hosted in the EU"
description: "Run Google's Gemma 4 26B A4B on inference.eu: a fast mixture-of-experts model with image input and 256K context, processed in EU data centres."
canonical: "https://inference.eu/models/gemma-4-26b-a4b/"
language: "en"
---

Vision · Google DeepMind

# Run Gemma 4 26B A4B in the EU through an OpenAI-compatible API

Gemma 4 26B A4B is the mixture-of-experts member of Google DeepMind's Gemma 4 family: about 25 billion parameters in total, but only 3.8 billion active per token. On inference.eu you call it with the OpenAI or Anthropic request format, your requests are processed in EU data centres, and nothing you send is used for training.

[Contact sales](https://inference.eu/contact/)[Example request](#example)

1.  0125.2B · 3.8B
2.  02EU data centres
3.  03Apache 2.0

## What Gemma 4 26B A4B is

Gemma 4 26B A4B is an open mixture-of-experts model from Google DeepMind, released in April 2026 under the Apache 2.0 licence. The model card lists 25.2 billion total parameters and 3.8 billion active parameters, a vision encoder of about 550 million parameters, and a context window of 256K tokens.

The A in the name stands for active parameters: only a small subset of the network works on each token, so the model runs much faster than its total size suggests. Like the rest of Gemma 4, it takes text and images in any order and produces text, and it is designed for reasoning, coding and agent workflows.

## What it does well, and where it falls short

This is a good fit for high-volume workloads where speed and throughput matter: classification, extraction, summaries, chat and image questions at scale. It reads screenshots, charts and scanned pages, supports tool calls, and keeps the full 256K context of the larger Gemma 4 model.

With under 4 billion active parameters, it is less thorough than the dense Gemma 4 31B on hard reasoning, long writing and complex code. It does not accept audio. For tasks where a wrong answer is expensive, test it against the 31B model or a larger reasoning model before choosing; for everyday tasks it is often good enough at a fraction of the compute.

## Calling Gemma 4 26B A4B through the API

inference.eu uses the OpenAI request format, so your OpenAI SDK code, LangChain, LlamaIndex and the Vercel AI SDK work unchanged. Set the base URL to the inference.eu API, use your inference.eu key and set the model to `gemma-4-26b-a4b`.

Send images as `image_url` content parts next to your text. Tool calls, JSON output and streaming work as usual. The Anthropic Messages format is also available, so tools built for Claude can use Gemma 4 26B A4B too. Because the request format is the same for every model, you can A/B test it against Gemma 4 31B by changing only the model string.

## Data residency for Gemma 4 26B A4B

On inference.eu, Gemma 4 26B A4B runs on infrastructure in the European Union. Your prompts, images and responses are processed in EU data centres, are not shared with Google, and are never used to train any model.

We sign a data processing agreement under Article 28 of the GDPR with every customer and provide our subprocessor list, so your records of processing can name every party involved. Billing is in euros, on the same invoice as every other model you use. The Apache 2.0 licence allows commercial use without user limits, so you can build the model into paid products straight away.

## Gemma 4 26B A4B

Maker

Google DeepMind

Made in

United States

Type

Vision

Parameters

25.2B

Active parameters

3.8B

Context window

262,144 tokens

Licence

[Apache 2.0](https://ai.google.dev/gemma/docs/gemma_4_license)

Released

April 2026

Model ID

gemma-4-26b-a4b

Processing

EU data centres

Training on your data

Never

Source

[huggingface.co](https://huggingface.co/google/gemma-4-26B-A4B-it)

gemma-4-26b-a4b · Python

```
import os
from openai import OpenAI

client = OpenAI(base_url="https://api.inference.eu/v1", api_key=os.environ["INFERENCE_EU_API_KEY"])

response = client.chat.completions.create(
    model="gemma-4-26b-a4b",
    messages=[{"role": "user", "content": "Summarise this contract in five bullet points."}],
)
print(response.choices[0].message.content)
```

## What teams use it for

-   1

    ### High-volume classification

    Tag support tickets, reviews or emails by topic, urgency and language across large daily volumes.

-   2

    ### Screenshot triage

    Read error screenshots and photos from users and route them to the right team with a short summary.

-   3

    ### Lightweight agents

    Run tool-calling agents for routine tasks such as lookups and form filling where speed matters more than depth.

## Gemma 4 26B A4B: common questions

The questions buyers, developers and legal teams ask us most. Something missing? Ask us directly.

That depends on how you process data. On inference.eu, requests to Gemma 4 26B A4B are processed in EU data centres, not used for training and covered by a data processing agreement under Article 28 GDPR with a published subprocessor list. That gives you a solid contractual basis for personal data.

A4B refers to roughly 4 billion active parameters. The model has 25.2 billion parameters in total, split into experts, but only about 3.8 billion of them are used for each token. That is why it generates text much faster than a dense model of the same size.

Gemma 4 26B A4B supports 256K tokens of context, the same as Gemma 4 31B. That is enough for long documents and multi-turn conversations. Images take up part of the context, so leave room for them when you send many pages.

Yes. Gemma 4 is released under the Apache 2.0 licence, which allows commercial use, modification and redistribution without user limits or extra usage terms. You can build it into paid products without a separate agreement with Google.

Choose 26B A4B for speed and volume: it uses about an eighth of the active parameters of the 31B model. Choose 31B when answer quality on harder reasoning, writing or coding matters most. Both share the same context window and image support.

Yes. It accepts images of different sizes and aspect ratios mixed with text in one prompt, and answers in text. It can read charts, screenshots and scanned documents. It does not accept audio; in the Gemma 4 family that is limited to the E2B, E4B and 12B models.

Point your OpenAI client's base URL at the inference.eu API, use your inference.eu key and set the model to \`gemma-4-26b-a4b\`. Messages, images, tool calls and streaming work with the same code you already have.

Both are small-active mixture-of-experts models with image input and 256K context. Qwen3.6 35B A3B is tuned towards agentic coding and has a thinking mode. Gemma 4 26B A4B is a general-purpose model from Google. Test both on a sample of your real traffic.

No. Requests to inference.eu are used only to produce your response. They are never used to train or fine-tune Gemma or any other model, and Google, which publishes the weights, has no access to them.

Google states support for more than 35 languages out of the box and pre-training on more than 140. Major European languages such as German, French, Spanish, Italian and Dutch are well covered. Check smaller languages against your own content before going live.

[Ask a question](https://inference.eu/contact/)

## Related

[All models](https://inference.eu/models/#directory)

-   [Gemma 4 31BRun Google's Gemma 4 31B on inference.eu: an open multimodal model with 256K context behind an OpenAI-compatible API, processed in EU data centres.Vision](https://inference.eu/models/gemma-4-31b/)
-   [Qwen3.6 35B A3BRun Qwen3.6 35B A3B on inference.eu: a fast mixture-of-experts model with 3B active parameters, image input and 256K context, processed in the EU.Code](https://inference.eu/models/qwen3-6-35b-a3b/)
-   [gpt-oss-20bRun OpenAI's compact open-weight gpt-oss-20b reasoning model on inference.eu: the OpenAI API format, processed in EU data centres, billed in euros.Reasoning](https://inference.eu/models/gpt-oss-20b/)
-   [Mistral Small 4 119BRun Mistral Small 4 119B from French lab Mistral AI on inference.eu: reasoning, vision and code in one model, processed in EU data centres.Chat](https://inference.eu/models/mistral-small-4-119b/)

## Move your AI workloads to EU infrastructure

Tell us which models you use and roughly how much traffic you have. We reply within one working day with pricing and next steps.

-   Reply within one working day
-   Pricing for your volume
-   DPA ready to sign