Gemma 4 Pushes Open Models Forward

Key insights

• Google wants open models to feel ready for real products, not just fun to download, by focusing Gemma 4 on hardware developers already own.
• The Apache 2.0 license may matter more than the model sizes because it makes Gemma easier for companies to use in commercial products.
• Gemma 4's lineup suggests Google wants one open model family that can scale from phones and connected devices to desktops and local coding setups.
• Planning, long context and tool use are quickly becoming expected features even in smaller open model launches.

In Brief

Olivier Lacombe introduces Gemma 4 as Google's newest family of open models, built from the same research base as Gemini and designed to run directly on hardware developers already own. The two biggest messages in the launch are practical: Gemma now uses the Apache 2.0 license, and the family is explicitly aimed at multi-step AI tasks on local devices.

That combination matters more than the short demo format suggests. Google is not only refreshing model sizes. It is trying to make open models easier to use in real products, from phones and connected devices to desktops running local reasoning and coding tools.

What changed in Gemma 4

The headline product changes are straightforward. Google says Gemma 4 is released under an Apache 2.0 license for the first time, built for complex logic and multi-step AI tasks, and designed to run on devices people already control. That is a stronger signal for real product use than earlier "open model" branding on its own.

The larger models include a 26B mixture-of-experts model and a 31B dense model. Google positions them as local reasoning and coding models for personal computers, with the bigger setup supporting up to a quarter million tokens of context and built-in tool use. That is a direct pitch at developers building assistants that need to plan, fetch information and take actions.

At the smaller end, Google highlights 2B and 4B models designed for memory efficiency on mobile and connected devices. The company also claims combined audio and vision support for real-time processing and support for more than one hundred forty languages.

---

Why the license change matters

The Apache 2.0 detail may be the most strategically important statement in the whole video. Open models are more useful when companies can adopt them without getting stuck in legal and product approval questions. Google's decision to stress the new license suggests it knows the bottleneck is not only model quality. It is whether teams can actually use the model inside real products.

That fits the rest of the launch framing. Gemma 4 is presented less as a research curiosity and more as a model family for developers who want local control. Lacombe explicitly pitches local reasoning and coding without needing to upload data outside a controlled environment. Google presents Gemma 4 as models that can be downloaded and run on hardware you control. That opens the door to setups where both the model and the data can stay local, without content needing to be sent to external cloud services. In practice, that is as much a privacy and system-design argument as it is a model argument.

---

Why this matters for the open model market

Gemma 4 shows how launch language around AI models has changed. Reasoning, long context, tool use and multi-step behavior are no longer niche features. They are becoming baseline expectations. Even a two-minute product overview now assumes developers want models that can do more than answer simple text requests.

Google is also making a broader market bet: that open models will matter across a full device ladder, not just in cloud or workstation environments. If one family can span phones, laptops, desktops and connected hardware, then open local AI becomes less of a hobbyist setup and more of a product strategy.

The limitation is that this video is still a launch summary, not a deep technical comparison. It tells you what Google wants Gemma 4 to mean. It does not fully prove how Gemma 4 performs against rivals in real production use. But as positioning, the message is clear: Google wants Gemma to be the open model line developers can actually build around.

---

Practical implications

• If you build local AI products, Gemma 4 is more relevant than earlier Gemma releases. Apache 2.0 and tool use reduce friction for real deployment.
• If you care about privacy or controlled environments, the local-compute pitch is central. Google is explicitly selling Gemma 4 as a way to keep work on hardware you own.
• If you compare open models, look beyond size. The meaningful differentiators here are license, context, tool use and where the models are supposed to run.

---

Glossary

---

Sources and resources