NEW Gemma 4 Update!

Key Takeaways

New Gemma 4 update. You've been paying for AI tools you don't need to pay for. You've been waiting for cloud-based models when you could run one locally. You've been limited by tools that can't work offline, can't run on your phone, and can't be customized. And you might have missed one of the biggest open AI releases of the year. Google just dropped Gemma 4, and this one is different. Stay with me because I'm going to break down every model, every feature, and exactly how you can start using this today. I'm the digital avatar of Julian Goldie, and if you're here, you want to actually understand AI tools and use them in real work, not hype, not theory, real practical stuff. Let's get into it. Gemma 4 just launched. This is Google's most capable open model family ever built, and I want to be really clear about why this matters because there's a lot of noise out there, and most of it misses the point. Here's what's actually happening. Google has been building a family of open models called Gemma since early 2024. The community went wild. Developers downloaded those models over 400 million times. People built more than 100,000 custom versions on top of them. That's a Gemma ecosystem. It's huge. And now Google just released the next generation. Gemma 4 is built on the same research and architecture as Gemini 3, Google's most advanced proprietary model. You're getting frontier-level AI research baked into a model you can run yourself on your own hardware, completely free. Let me walk you through what's actually new because there's a lot to cover. First, let's talk about what Google calls intelligence per parameter. This sounds technical, but it's actually really simple. Normally, bigger models are smarter. That's just how it works. If you want better output, you run a bigger model. But Gemma 4 changes that. Google has figured out how to pack more intelligence into fewer parameters. What that means for you is that a smaller model can now do things that previously only a much larger model could do. You get better results with less hardware, less memory, less power, less cost. The 31B dense model in the Gemma 4 family currently ranks third among all open models on the Arena AI text leaderboard, and that's a widely used independent benchmark. Third place across all open models. That's not a marketing claim. That's a measured ranking. And this is running on hardware regular people can actually afford. Now, there's a second thing coming up that a lot of people miss entirely. I'll get to it in a minute. Let me break down the four models in the Gemma 4 family because each one is built for a different use case, and picking the right one matters. The first two are called E2B and E4B. The E stands for effective. These are the small models. They activate an effective 2 billion and 4 billion parameter footprint during inference. What that means practically is they're designed to run on mobile phones, tablets, Raspberry Pi, IoT devices, anything low power. And they don't just handle text, they handle images, video, and audio, native audio, on device, no internet needed. The E2B model can run with under 1.5 GB of memory on some devices thanks to 2-bit and 4-bit weight compression. If you've tried running AI locally before and hit a wall with RAM, that number should mean something to you. These two models also support over 140 languages. So this is not a model built only for English speakers. This is built for global deployment. The other two models are the 26B mixture of experts and the 31B dense. These are the bigger, more powerful options. They're designed for personal largest one can run on a single 80 GB H100 GPU. That's what Google calls frontier-level capability without requiring enterprise infrastructure. Both of the larger models support a 256,000 token context window. That means you can feed in an entire code base, a full set of documents, a massive project, all in one prompt. The smaller models support 128,000 tokens, still very large. Now, here's the thing I said I'd come back to, agentic AI. Most models need you to fine-tune or hack them to work as agents, to call tools, to take actions, to handle multi-step workflows. Gemma 4 has native support for function calling and structured JSON outputs built in from the start. No workarounds needed. This means you can build an AI agent that interacts with third-party software, executes a plan across multiple steps, and works entirely offline. That's a big deal for anyone building automation workflows or deploying AI in environments where internet access is limited or restricted. Google already has agent skills running on device through the Google AI Edge Gallery app. These agents can access external knowledge, summarize content, turn documents into flashcards, create visual data representations, all running locally on your phone. I was honestly surprised by how practical this is right now. And if you're wondering how to actually implement this in your own work, I'll come back to that. But first, there's one more piece you need to know about. Let me talk about the license for a second because this is actually important. Gemma 4 is released under Apache 2.0. That is about as open as a license gets. You can use it commercially. You can modify it. You can build on it and distribute what you build. There are no hidden restrictions here. Google heard feedback from the developer community about previous licensing limitations, and they changed it. This is a model family you can genuinely build a product on without worrying about licensing headaches. Compare that to some other model releases where commercial use requires a separate agreement, or there are restrictions on how many users you can serve, or you have to check back with the company. With Apache 2.0, none of that applies. When I first started digging into local AI models and trying to figure out which ones were actually worth deploying in real workflows, I was completely overwhelmed. There are so many models, so many benchmarks, so much noise about what's best. That's when I created a community called AI Profit Boardroom. We have over 2,000 members all focused on learning AI together and sharing what actually works. It taught me which workflows save time versus which ones waste it. The community shares real use cases and practical implementations. If you're serious about using AI to improve your work and skills, check it out. Link in description. Back to Gemma 4. Let's talk about where you can actually get it and what you need to run it. For the smaller E2B and E4B models, you can try them right now in the Google AI Edge Gallery app. It's available on iOS and Android. You download the app, download the model, and you're running local AI on your phone. Done. For the larger 26B and 31B models, you can access them in Google AI Studio today. If you want to run them locally, you can download the weights from Hugging Face, Kaggle, or Ollama. And here's something that matters. Gemma 4 has day-one support for a huge list of tools. Hugging Face Transformers, VLM, Yamas Ollama, LM Studio, Nvidia NIM, Docker. The list goes on. So whatever your current setup is, there's almost certainly a way to plug Gemma 4 directly into it without rebuilding your stack. It's also optimized for Nvidia GPUs, AMD GPUs through the ROCm stack, and Google Cloud TPUs. So whether you're on consumer hardware or running on cloud infrastructure, it's covered. For Android developers specifically, you can now prototype agentic workflows using the AI Core Developer Preview, and Google has said it's forward-compatible with Gemini Nano 4. So you're building on something that has a clear roadmap forward. Let me give you some real examples of how people are already using this. The community has already built over 100,000 custom variants on previous Gemma models. One example Google specifically mentioned is Yale University's cell-to-sentence scale model, which researchers built on top of Gemma for cancer research. And there's also a Bulgarian first language model someone built specifically optimized for that language. That's what open models unlock, specialized, customized, domain-specific AI that a commercial API would never offer because there's no business case for it. But researchers, developers, and people with specific needs can build exactly what they need. Let me zoom out for a second and put this in context. There is a real competition happening in the open model space right now. Meta has Llama, Alibaba has Qwen, Mistral is building, Deep Seek made waves earlier this year, and now Google is pushing hard with Gemma 4. This is good for everyone. More competition means better models. Better models means you get more capable tools at lower costs. And the open-source nature of all of this means you're not locked into any single provider. What makes Gemma 4 particularly interesting is the combination of the edge focus, the agentic capabilities baked in natively, the Apache 2.0 license, and the fact that it's built directly from Gemini 3 research. That's not a common combination. Usually, you have to pick either you get a very capable model that requires serious hardware, or you get a small efficient model that has limited capabilities. Gemma 4 is trying to give you both depending on which variant you pick. If you want to start today, here's exactly what I suggest. If you have an Android or iOS phone, download the Google AI Edge Gallery app and play with E2B or E4B. See what on-device agentic AI actually feels like. It's free. It runs offline. It's genuinely impressive for what it is. If you have a decent computer and you're already running local models, grab the 27B or 31B from Hugging Face or Ollama and run it through your existing setup. The day-one tool compatibility means you probably don't need to change anything. If you're a developer building products, the Apache 2.0 license means you can build on this commercially without restrictions. The native function calling and JSON support means agentic workflows are easier to build than they've ever been on an open model. Gemma 4 is available right now. It's free. It's open. And it's genuinely one of the more capable open model releases we've seen. I'll be covering more on this as the community builds more on top of it. There are already use cases being developed that I want to dig into more deeply. So if you're not subscribed yet, that's your next move. If you're looking to dive deeper into AI tools and actually implement them in your work, I recommend AI Profit Boardroom. Over 2,000 people learning how to use AI effectively. Everyone shares real experiences, what's working, what's not, which tools are worth your time, which ones to skip. No hype, just solid information and practical guidance from people doing the work. It's helped me stay on top of updates and figure out how to actually apply them. Link in description if you want to check it out. And if you want the full process, SOPs, and 100-plus AI use cases like this one, join the AI Success Lab. Links in the comments and description. You'll get all the video notes from there, plus access to our community of 58,000 members who are crushing it with AI. See you in the next one.