Gemma 4 First Look: Google's Open Model for Agents You Run Yourself

Launched: April 2026 · Reviewed: 14 April 2026 · Score: 18/25 ⭐⭐⭐⭐ Very Good

Verdict in one line: The most approachable serious open model of autumn 2026 — Google quality, no subscription, and your data stays wherever you put the hardware.

What launched

In April 2026 Google released Gemma 4, the next generation of its open-weights model family — the downloadable cousins of Gemini, which have passed 500 million downloads as a series. Gemma 4's pitch is advanced reasoning and agentic workflows in an open model: tool use and multi-step task execution, not just chat.

What it does

Gemma 4 is a model you possess rather than rent: download the weights, run them on your own hardware or cloud, fine-tune on your own documents. The agentic angle matters — earlier small open models chatted well but fell apart orchestrating tools; Gemma 4 is built for exactly the workflows businesses now want to automate privately.

Pricing for Australians

Free to download and use, including commercially, under Google's Gemma licence — note it is not Apache 2.0; the Gemma terms carry use restrictions worth reading before you build a product on it. Running costs are your hardware or cloud bill; smaller variants typically run on consumer GPUs, which has been the Gemma family's calling card.

Australian Fit

Like Mistral Small 4 (see our March first look), the fit case is sovereignty: run it onshore and client data never leaves the country — gold for legal, accounting and health practices under the Privacy Act. Strong local tooling support in Ollama and LM Studio makes Gemma models the usual first download for Australian local-AI beginners. 4/5.

Early read

Good: the Gemma family's small-hardware heritage plus agent capability is a practical combination, not a benchmark stunt; ecosystem support arrives day one because every local-AI tool ships Gemma configs. Not so good: the bespoke licence, not Apache 2.0, adds legal reading; open models still trail their Gemini siblings on hard tasks; and "free model" never means "free system" — someone has to run, update and secure it.

Early verdict

If you've been waiting for a reason to try local AI for your business, Gemma 4 in Ollama or LM Studio is the gentlest serious starting point. If you're already running Mistral Small 4, benchmark both on your actual workload — the answer differs by task more than the leaderboards suggest.

First-look score — provisional until our full hands-on review.

Frequently asked questions

Is Gemma 4 available in Australia?

Yes. Gemma 4 is an open-weights model you download and run yourself, so it's available anywhere — Australia included. Strong local tooling support in Ollama and LM Studio makes it the usual first download for Australian local-AI beginners.

Does Gemma 4 cost anything?

It's free to download and use, including commercially, under Google's Gemma licence. Note the licence is not Apache 2.0 — the Gemma terms carry use restrictions worth reading before you build a product on it. Your only running cost is your own hardware or cloud bill.

Why is Gemma 4 a good fit for Australians?

Because you can run it onshore, so client data never leaves the country — valuable for legal, accounting and health practices under the Privacy Act. It's the gentlest serious starting point if you've been waiting to try local AI.

What are the catches?

The bespoke (non-Apache) licence adds legal reading, open models still trail their Gemini siblings on hard tasks, and "free model" never means "free system" — someone has to run, update and secure it.

Read next

• Category guide: Best Local & Private AI Tools in Australia 2026
• AnythingLLM Review Australia 2026
• GPT4All Review Australia 2026

Sources

• Google — April 2026 AI updates (Gemma 4): https://blog.google/innovation-and-ai/technology/ai/google-ai-updates-april-2026/ (primary)
• Google AI for Developers — Gemma: https://ai.google.dev/gemma (primary)