Gemma 4 Is Here: Google's Open-Source AI Runs Locally, Builds Apps, and Doesn't Need the Cloud

What It Actually Does When You Test It

Benchmarks are fine. What I care about is what happens when you hand the model a real task and walk away. One YouTuber who covers AI coding tools ran the 31B model through a series of front-end generation challenges using the Kilo CLI as the agentic harness. First up: build a macOS-style operating system in the browser. The output had a working toolbar, app icons, a functional calculator, a terminal, and a settings panel — all styled correctly. Not everything was interactive, and the SVG icons were a bit rough, but for a 31B parameter model generating full UI code in one pass, it's rated an 8 out of 10 [1]. Then the Airbnb clone test. The model reproduced the layout convincingly — icons, cards, formatting, navigation structure. It looks like the actual Airbnb website. Then an F1 donut simulator with physics-like motion. Then a product viewer with 360-degree rotation, color-switching, and hotspot annotations. Each of these was a single agentic run with no manual corrections between steps [1]. The 26B model — the one you can actually run locally on a MacBook Pro or similar hardware — kept pace in most of these tasks. The outputs are messier: some animations break, some components don't fully wire up. But the fundamental structure is correct, and you can iterate from there. If you're a developer using this as a starting point rather than a finished product, the 26B is legitimately useful on-device [1]. One comparison that's worth flagging: the 31B uses roughly 2.5 times fewer output tokens than Qwen 3.527B for similar tasks. That means faster generation and lower API costs, even if the Qwen 3.5 scores slightly higher on raw intelligence benchmarks. In real-world use, token efficiency matters as much as capability — especially if you're running this at scale [1].

The Agentic Play — Agents That Run On Your Phone

Here's the part that's genuinely new territory: alongside Gemma 4, Google rolled out something called "agent skills" — a framework for running multi-step agentic tasks entirely on-device through the Gemini app. In the demos, the model doesn't just answer questions. It decides which tools to use, in what order, and how to combine the outputs. You can ask it to pull structured data from your phone, process it, and generate a visualization — all in one flow, with no internet connection required. The entire agent loop runs locally [1]. For most people, this sounds abstract. Here's the concrete version: imagine you're on a plane with no Wi-Fi and a 500-page legal contract PDF you need to summarize before you land. Old AI: sorry, needs the internet. Gemma 4: the phone handles it [2]. Multimodal means it can process images and audio on top of text — show it a photo of your fridge, it suggests a recipe. Voice input is native. For developers building mobile apps or offline-capable tools, this is a meaningful architectural shift. Cloud inference has dominated AI app development because local models weren't good enough. That's changing fast. The 4B model running agent skills on a phone is a different world than what local models could do 12 months ago.

How to Run It

The weights are available through Hugging Face, Ollama, and LM Studio. If you want to test the API before committing to local setup, Google AI Studio has it free to use. The 31B via API runs at 14 cents per million input tokens and 40 cents per million output tokens — cheap for cloud inference [1]. The community is also recommending the Kilo CLI as the agentic harness of choice, which provides $25 in free credits and surfaces the model's tool-use capabilities more naturally than plain API calls. If you want to test what Gemma 4 can do with actual agentic tasks rather than just chat completions, that's the right setup.

Why This Matters More Than Another Big Model Drop

Every few weeks there's a new "biggest model" announcement. Most of them are cloud-only, closed-weight, and irrelevant to anyone who isn't paying enterprise API bills. Gemma 4 is different because it's pointing at a different future: AI that runs on the hardware you already own, doesn't send your data anywhere, and costs you nothing beyond compute. The 26B model runs at real-world speeds on a mid-range workstation. The 4B and 2B models are designed for phones and edge devices. There's an honest read on Google's motivation here — Apache 2.0 licensing is a developer acquisition play. Get every dev in the world building on Google's tools, sell them cloud compute later. That's not a secret [2]. But the trade-off for users and developers is genuinely good: a model family that performs competitively, runs locally, costs nothing to license, and is improving fast. The open-source AI gap with proprietary models is narrowing faster than the closed-model labs want to admit. Gemma 4 is part of that story — a 31B model ranking third on LM Arena is not where open-source was a year ago. If the trend holds, the question won't be "why would I run a local model" but "why would I pay for cloud inference when the local model is good enough?" We're not at that crossover yet. But Gemma 4 closed the gap.