Name: ANT
Author: Antbase

Google DeepMind just released Gemma 4 — a family of four open-weight models built from the same research behind Gemini 3. They ship under Apache 2.0, which means no usage caps, no acceptable-use gatekeeping, full commercial freedom. And they're genuinely impressive.

Why Gemma 4 Matters

The headline number: Gemma 4 31B ranks #3 globally on Arena AI's text leaderboard with a 1452 Elo — outperforming models twenty times its size. It scores 74.4% on BigBench Extra Hard. For context, that puts a 31-billion-parameter model in the same tier as the flagship closed-source models from just a few months ago.

But raw benchmarks only tell part of the story. Every Gemma 4 variant — even the tiny 2B — supports multimodal input (images, video, audio), native function calling, structured JSON output, multi-step planning, and configurable extended thinking. These aren't stripped-down distillations. They're full-featured models that happen to be small enough to run on a phone.

The Four Models

•Gemma 4 E2B — 2.3B effective parameters, 128K context. Runs on smartphones and Raspberry Pi. Multimodal.
•Gemma 4 E4B — 4.5B effective parameters, 128K context. Higher reasoning quality, same tiny footprint.
•Gemma 4 26B MoE — 25.2B total params but only 3.8B active per forward pass. Nearly as fast as a 4B model with 256K context.
•Gemma 4 31B Dense — The flagship. 256K context. Arena Elo 1452. The one you want for serious work.

The MoE variant is the sleeper hit. With only 3.8 billion parameters firing per token, it runs at roughly 4B speed but delivers 26B quality. If you're cost-sensitive or latency-sensitive, this is the one to watch.

Apache 2.0 — Actually Open

Previous Gemma releases shipped with Google's custom license that had commercial restrictions. Gemma 4 switches to Apache 2.0 — the same license as Qwen 3.5, and more permissive than Llama 4's community license. No monthly active user limits, no policy enforcement, no lawyer needed. Fork it, fine-tune it, ship it, sell it.

Using Gemma 4 on Antbase

Gemma 4 models are already available on Antbase through multiple providers. ANT's router automatically picks the fastest available endpoint, so you don't need to think about which provider is hosting it — just use the model name.

python

from openai import OpenAI

client = OpenAI(
    base_url="https://api.antbase.ai/v1",
    api_key="your-ant-api-key",
)

response = client.chat.completions.create(
    model="google/gemma-4-27b-it",
    messages=[
        {"role": "user", "content": "Explain quantum entanglement like I'm five."}
    ],
)
print(response.choices[0].message.content)

typescript

import OpenAI from "openai";

const client = new OpenAI({
  baseURL: "https://api.antbase.ai/v1",
  apiKey: "your-ant-api-key",
});

const response = await client.chat.completions.create({
  model: "google/gemma-4-27b-it",
  messages: [
    { role: "user", content: "Write a haiku about distributed systems." },
  ],
});
console.log(response.choices[0].message.content);

You can also use ant:free to let the router automatically select Gemma 4 when it's the best free option for your request. No configuration needed — if Gemma 4 is the strongest free model for your prompt, ANT will route to it.

Try It in the Playground

The fastest way to test Gemma 4 is the Antbase Playground. Select the model from the dropdown, type a prompt, and see the response stream in real time. No API key required for the playground — just log in and start chatting.

Try Gemma 4 27B Try Free Tier (Auto)Try Auto Routing

Extended Thinking

Gemma 4 supports configurable extended thinking — the model can reason through complex problems step by step before producing its final answer. On Antbase, just set reasoning_effort in your request and ANT handles the rest, whether the request lands on Gemma 4, Claude, or any other reasoning-capable model.

python

response = client.chat.completions.create(
    model="google/gemma-4-27b-it",
    messages=[
        {"role": "user", "content": "Prove that the square root of 2 is irrational."}
    ],
    extra_body={"reasoning_effort": "high"},
)

Who Should Use Gemma 4

•Developers building AI products who need a strong open model with no licensing headaches
•Teams running on-prem or in regulated environments where data can't leave the network
•Anyone who wants flagship-tier quality without flagship-tier pricing
•Hobbyists and researchers — the small variants run on consumer hardware

The Bottom Line

Gemma 4 is the most capable open-weight model family released to date. The 31B dense model competes with models that cost orders of magnitude more to run. The MoE variant gives you 26B brains at 4B speed. And it's all Apache 2.0. Google just raised the floor for what open models can do — and you can try it right now on Antbase, for free.

Try Try Gemma 4 27B Try Free Tier