Gemma 4 free API: providers and limits

Where Gemma 4 is free

Google publishes the weights under Apache 2.0, and five providers serve the 31B or 26B A4B model on a free tier.

Provider	Model id	Free-tier terms
OpenRouter	`google/gemma-4-31b-it:free`	20 requests per minute, 50 or 1,000 per day. The 26B A4B model is free too, as google/gemma-4-26b-a4b-it:free.
Gemini	`gemma-4-31b-it`	Free tier only (no paid tier for Gemma); also gemma-4-26b-a4b-it. Limits are shown per project in AI Studio. Prompts may be used to improve Google products.
SambaNova	`gemma-4-31B-it`	Free tier: 20 requests per minute and 20 per day. Preview model, 128K context.
Requesty	`google/gemma-4-31b-it`	200 requests a day and 20 a minute for new organizations, shared across all free models.
Cloudflare	`@cf/google/gemma-4-26b-a4b-it`	The 26B A4B model, within the shared 10,000 Neurons a day.

Listed, but not free

NVIDIA (google/gemma-4-31b-it): trial credits for prototyping, not a free tier.

What Gemma 4 is

Gemma 4 is Google DeepMind's family of open-weight multimodal models, dense and mixture-of-experts, sized from phones to workstations for reasoning, coding, agentic and multimodal work. The two hosted sizes are the 31B dense model and the 26B A4B mixture-of-experts.

Hosted sizes: 31B dense; 26B A4B MoE (25.2B total, 3.8B active)
Other sizes: E2B and E4B for phones, 12B for workstations
Context window: 256K tokens (31B and 26B A4B)
Input: Text and image
Output: Text
Tools and reasoning: Native function calling, configurable thinking

Before you rely on it

The free Gemma 4 models on Routeway are community fine-tunes of 26B A4B, not Google's model.
On the Gemini API, apps offered to users in the EEA, Switzerland or the UK must use paid services, and Gemma 4 has no paid tier there.
SambaNova serves 31B with a 128K context, half the native 256K.

Use Gemma 4 with automatic failover

A single free tier runs out. Ficelle puts the free models of every provider you configure behind one local OpenAI-compatible endpoint. Call a routing profile and Gemma 4 is one of the models it can pick; when it rate-limits or fails, the next eligible free model answers. To pin Gemma 4 alone, call the id that ficelle models lists for it, knowing a pinned model has no failover.

python

from openai import OpenAI

client = OpenAI(base_url="http://127.0.0.1:8646/v1", api_key="ficelle-local")

# A routing profile keeps failover: Gemma 4 is one of the free
# models it can pick, and the next eligible one answers if it fails.
client.chat.completions.create(
    model="ficelle/auto-tools",
    messages=[{"role": "user", "content": "..."}],
)

Routing profiles that fit what Gemma 4 is built for:

Whether a profile routes to Gemma 4 on your install depends on your providers' live catalogs and Ficelle's eligibility checks, decided at request time.

Questions

Is Gemma 4 free to use?

Google publishes the weights under Apache 2.0, and five providers serve the 31B or 26B A4B model on a free tier. When we checked on 28 September 2026, OpenRouter, Gemini, SambaNova, Requesty and Cloudflare served it at no charge, each under its own free-tier limits.

Where can I use the Gemma 4 API for free?

OpenRouter as google/gemma-4-31b-it:free; Gemini as gemma-4-31b-it; SambaNova as gemma-4-31B-it; Requesty as google/gemma-4-31b-it; Cloudflare as @cf/google/gemma-4-26b-a4b-it. Free catalogs change often, so check the provider's own page before you rely on one.

What is Gemma 4?

Gemma 4 is Google DeepMind's family of open-weight multimodal models, dense and mixture-of-experts, sized from phones to workstations for reasoning, coding, agentic and multimodal work. The two hosted sizes are the 31B dense model and the 26B A4B mixture-of-experts.

Can I call Gemma 4 through Ficelle?

Yes, when one of your configured providers serves it for free. Run ficelle models to see the ids your install can route to. Calling that id pins the one model, with no failover; calling a routing profile such as ficelle/auto-tools keeps failover, so another free model answers when Gemma 4 is rate-limited.

Sources

Official pages this guide rests on, read on 28 September 2026: