Gemma 4 Just Changed Open Source AI Forever

Jemma 4 just changed open source AI forever, so Google just dropped Jemma 4 and it runs on your phone, offline for free. Let that sit for a second because what just happened here is bigger than most people realise. The Jemma series has been downloaded over 400 million times. Developers have built more than 100,000 variants on top of it, and now Google just released the most powerful version yet, built from the same research as Jemma 9.3, their flagship commercial model, and gave it away, free, open, Apache 2.0 licensed, no restrictions, no legal asterisks, no calling your lawyer before you use it.

This is not a small thing, this is Google saying we're done playing defence in the open source AI space. So there's four models and one for every device you own. Jemma 4 comes in four sizes and here's why that matters. The two smaller ones, E2B and E4B, are built for phones, laptops and edge devices.

E2B runs with under 1.5 gigabytes of memory. The E4B needs around 1.5 gigabytes. Both handle text, images and audio. Both have a 128k context window.

Both run completely for free offline. Now think about what that means for a second. You've got a multimodal AI model, one that can read documents, analyse images, understand audio and it runs locally on your phone. No internet connection, no API costs, no data going to any server anywhere.

A local marketing agency, for example, could run Jemma 4 E4B directly on a laptop, process client briefs, analyse competitor apps and generate copy all without a single API call or monthly subscription. And then there's the bigger two. So you've got the 26B MOE, mixture of experts, which activates only 3.8 billion parameters during inference. So you get near flagship performance at a fraction of the compute cost.

The 31B dense model is the powerhouse. It fits on a single 80 gigabyte H100 GPU, scored 1,452 on the LM Arena leaderboard and sits at number three globally at the time of release. Number three, open source, free. So why is Apache 2.0 the real story here?

Well, for two years, businesses looking at Jemma had a problem. The model was good, but the license was complicated. Custom terms, usage restrictions, clauses that required legal review before you could deploy it commercially. So teams went with Quen, for example, or Mistral, or models with cleaner, simpler Apache 2.0 terms.

But Jemma 4 just removed that friction entirely. Apache 2.0 means you can use it, modify it, build products on it, sell those products with no restrictions, no usage caps, and no terms Google can update and surprise you with later. It's the same license that Quen uses, the same one that Mistral uses. And now Google is on that same playing field.

And their model just ranked third in the world. If you've been waiting for a reason to build on Google's open models, this is it. The legal team finally has nothing to flag. And here's what it can actually do.

And this is where it gets genuinely interesting for business owners. So Jemma 4 is multimodal. It handles text, images, and on the smaller models, audio. It has native function calling built in.

That means it can use tools, run workflows, take multi-step actions on its own. And it's not just a chat model. It's an agent foundation. It supports over 140 languages out of the box.

The larger models have a 256K token context window. That's roughly 200,000 words of context in one single prompt, long enough to process an entire business's worth of documents, contracts, email chains, et cetera, in a single pass. And it has a built in reasoning mode, right? So you can turn on extended thinking where the model works through a problem step-by-step before giving you an answer.

Or you can turn it off for speed. Your call. You can analyze video by processing sequences of frames. It can read charts, PDFs, handwritten notes, and UI screenshots.

It can output structured JSON directly. And it supports bounding box detection for screen elements, which means that it can navigate websites and apps autonomously. A solo operator, for example, running a content business could feed Jemma 4 a month of their own content, competitor content, and audience data all in one context window and get a full content strategy back. Now, Jemma 4 running free offline on any device with full agent capabilities is exactly the kind of shift that changes what's possible for business owners in 2026.

Inside the AI Profit Boarding, we've already got members mapping out how to use open models like Jemma 4 to cut AI costs to zero whilst running full automation workflows. We've got 30 day roadmaps for setting up local AI systems, daily tutorials walking you through the exact tools and setups, and four weekly coaching calls every week where we dig into this stuff live. 2,700 business owners, a lot of them already running open workflows for lead gen, client work, and content production, plus a prompt library built around agency workflows and a member map so you can find and connect with people near you doing the same sort of thing. Link in the comments description or go to the AIProfitBoarding.com to get access.

And here's the acceleration curve that nobody's talking about. So the thing that keeps standing out to me is this. A year ago, running a frontier AI model locally meant a high-end technical setup and a model that was still a generation behind the closed ones. Now you've got a model that ranked third in the world running on your laptop offline for free with multimodal support, native agents, 256k contacts, and a clean commercial license.

The gap between closed and open just got significantly smaller. And that gap is closing faster than most people expected. Quen, Mistral, Lama, the open source space has been heating up for 18 months, but what Google just did with Gemma 4 is totally different. This isn't a community-built model that punches above its weight.

This is Google DeepMind releasing Gemini 3 research as an open model. They're bringing the A game to the ecosystem, right? To the open ecosystem. The E4B Edge model, one of the smallest in the family, hit 42.5% on AIME 2026 math benchmarks and 52% on LiveCodeBench.

These are the small models, right? Running on a T4 GPU, outperforming Gemma 3's 27B model on most benchmarks, despite being a fraction of the size. The intelligence per parameter jump here is real. Here's what this means if you're running a business.

Let's be direct about what the shift means for you, because AI API costs have been a real barrier. Paper token pricing adds up faster if you're running automation at scale. Processing leads, generating content, analyzing data daily. Gemma 4 actually changes that maths completely.

You run it locally, you pay once for the hardware, and then it's done. No monthly bill, no per call pricing. And because it's Apache 2.0, you can build client-facing products on it. An agency, for example, could build a custom AI tool for their clients powered by Gemma 4 and actually charge for it with no licensing friction and no ongoing costs eating into margin.

The offline capability matters more than people are giving it credit for. And this could be used for sensitive client data, financial information, private business details, because none of that needs to leave your machine with Gemma 4. You just process it locally, you keep control. And Google has also partnered with Qualcomm, MediaTek, NVIDIA, and the Pixel team to make sure Gemma 4 works well on the hardware people actually own.

So this isn't theoretical on-device AI. It's optimized for real devices tested on real hardware. AMD actually rolled out full support across their GPUs and CPUs within 24 hours of launch. NVIDIA is distributing Gemma 4 through their RTX AI Garage.

It's available on Hugging Face, Alarma, LM Studio, Kaggle, Google AI Studio, basically everywhere you'd want to grab it from day one. Now let's talk about the open-source power shift here, because something worth paying attention to is while some AI Chinese labs have been pulling back from fully open releases on their newest models, Google just moved in the opposite direction. They opened up more with better models on better terms. That's a deliverable choice, right?

And it signals something. Google isn't treating open source as an afterthought. They're treating it as a competitive strategy, right? They want Gemma running on as many devices as possible, powering as many products as possible, with Google's infrastructure underneath it all when teams need to scale.

The 400 million downloads and 100,000 community variants from previous generations weren't an accident for Gemma, Google built that ecosystem intentionally, and now they're feeding it with their best research. The direction of travel here is clear. Open models are getting better, faster, and the closed ones are sort of edging in in terms of distance, right? Open source is catching up with closed models, and the licensing is finally clean enough that businesses can actually build on them without hesitation.

So what should you do right now? Well, if you're running any kind of business where AI is part of your workflow, or you've been wanting it to be, Gemma 4 is worth your attention this week. The smallest model runs on a phone. Pull up Google AI Edge Gallery and try the E2B or E4B today.

No setup, no API key. You can just see what a local AI agent feels like to use. If you've got a laptop or desktop with a decent GPU, grab the 26B MOE through Alarm Studio. It's fast, it's free, and it handles multi-step tasks, image analysis, and document processing out of the box.

If you're thinking about building something on top of it, a workflow, a tool, a client product, for example, the Apache 2.0 license means you can just start today, right? No friction, no legal review, just build. And if you want to know exactly how to integrate Gemma 4 into your business workflows, how to set it up, what to use it for, how to combine it with automation tools to save real time and generate more leads, come join us in the AI Profit Board. We've got members right now building local AI workflows using open models, cutting the AI costs completely, and using the money they're saving to reinvest into business growth.

Four coaching calls every week, daily step-by-step tutorials, a 30-day roadmap built around AI automation for business owners, and 2,700 members inside. Link in the comments description or go to theaiprofitboard.com to get access. Gemma 4 just made powerful AI free. The only question now is whether you use it.