On April 2, 2026, Google took a quiet but significant step forward: the release of Gemma 4.
As an independent iOS developer who also runs an art practice across different countries, the question of whether I can run AI on-device is anything but academic. Relying entirely on cloud APIs means dealing with latency, connectivity costs, and privacy tradeoffs. But lightweight on-device models have historically sacrificed too much accuracy to be truly useful. Gemma 4 looks like a serious attempt to close that gap.
Here's what I found after spending time with the new models.
What Gemma 4 Is: Four Sizes, One Big Shift
Gemma 4 is the latest release in Google's open model series, available in four carefully tuned sizes:
- Effective 2B — designed to run on smartphones and even Raspberry Pi
- Effective 4B — the most balanced size for edge deployment
- 26B Mixture of Experts (MoE) — built for high-performance cloud workloads
- 31B Dense — currently ranked #3 globally on the Arena AI text leaderboard
The most important change, though, is the license. Gemma 4 ships under Apache 2.0, which means unrestricted commercial use, modification, and redistribution. For individual developers and small teams, this removes a major barrier to building with open models.
What Surprised Me: Multimodal Support Across All Sizes
Gemma 4 supports text and image input at every size, with audio input added for the smaller edge models (2B and 4B). This matters more than it might initially sound.
One of the first things I tested was using Gemma 4 to assist with UI prototyping. I passed in a screenshot of an app interface and asked it to identify accessibility issues. It flagged color contrast ratios, tap target sizing, and text readability — all concisely and accurately. The same kind of analysis I would normally run through the Gemini API, but this time entirely local.
That's the real value proposition: when privacy matters or connectivity isn't guaranteed, processing on-device changes what your app can promise users.
The 256K token context window was another standout. Being able to load a large codebase or a lengthy conversation history without truncation opens up genuinely new use cases. I tested it by pasting in a set of log files and asking for a debugging summary — something that used to require careful chunking or a cloud model.
Benchmark Improvements Worth Paying Attention To
The 31B model's performance gains are substantial. On AIME (mathematics), scores jumped from 20.8% to 89.2%. On LiveCodeBench (coding), from 29.1% to 80.0%.
In practice, this translated to noticeably better code review quality. When I ran it over some Swift async patterns I had been writing, it caught a potential memory leak in a closure that I had overlooked. The fix it suggested was accurate and idiomatic. That's the kind of precision you previously associated with much larger cloud models.
The Gemini Nano Connection: Why This Matters for Android Developers
Google has confirmed that Gemma 4 will serve as the foundation for the next generation of Gemini Nano — the model embedded in Android devices. If that update rolls out broadly, it means the on-device AI features available through Android AICore will gain multimodal capabilities and significantly improved reasoning.
The Google Android team has already announced a Gemma 4 AICore Developer Preview for compatible devices. Given that the 2B and 4B models are designed to run efficiently on mid-range hardware, widespread adoption could come faster than expected.
For Android developers, this is worth tracking closely. The gap between what you can do with on-device AI versus cloud AI has been narrowing — Gemma 4 is the most concrete signal yet that on-device may become the default, not the exception.
When to Use Gemma 4 vs. the Gemini API
After testing both, here is how I think about the tradeoff:
Stick with the Gemini API when you need access to the most capable models (like Gemini 2.5 Pro), require real-time web grounding, or are building experiences where users are reliably connected and accuracy is the top priority.
Reach for Gemma 4 on-device when you are processing sensitive user data you do not want to send to a cloud, need to function in offline or low-connectivity environments, want to minimize API costs at scale, or are targeting ultra-low latency for interactive features.
For apps in health, finance, or anything involving personal data, on-device inference is not just a performance choice — it is a trust choice.
How to Get Started Right Now
Getting hands-on with Gemma 4 is straightforward.
Google AI Studio lets you test all Gemma 4 sizes directly in the browser, with code examples generated automatically. If you already have Gemini API code, the transition is minimal.
Hugging Face hosts all Gemma 4 model weights. With the transformers library, you can load and run a model in just a few lines of Python — locally or in Google Colab.
Vertex AI supports Gemma 4 for teams already in the Google Cloud ecosystem who want to fine-tune or deploy at enterprise scale.
What I Am Most Excited About Going Forward
The Apache 2.0 license is the change I expect to have the longest tail. It makes community fine-tuning commercially viable, which means we should expect a growing ecosystem of domain-specific Gemma 4 variants — models optimized for legal, medical, educational, and coding use cases, built by developers around the world.
That kind of specialization is where open models have historically lagged behind proprietary ones. With Gemma 4's performance baseline and open licensing, that gap starts to close.
Google's commitment to open models has felt inconsistent at times, but Gemma 4 reads like a genuine strategic bet: four sizes, multimodal from day one, commercially open, with hardware reach that extends to consumer devices. If you have not tried it yet, Google AI Studio is the easiest place to start. I think you will be pleasantly surprised.