Google Gemma 4 Just Changed How Builders Ship AI

0:00In this video, we're going to be going over a new open weight model that just dropped and not by any old company. We're talking Google Gemma 4. We're going to be talking about what Google actually shipped and why I think this is one of the most important open model releases of the year and why this matters more for builders than for people who just like tweeting benchmarks screenshots all day. Hey, I'm guilty of that, too. But, this actually makes a huge difference for your building workflow because this part that jumps out to me is not just another new model. You know, they put a lot of thought into this one. You know, maybe it scores two points higher on some chart, cool, and you move on like other open weight models, but this one is completely different and hits a whole new nerve. So, Google took Gemma 4 and pushed it

0:47real hard toward real-world builder use, local use, agent use, and on-device use. And they put it under Apache 2.0. That last part matters a lot. We'll get to that in a bit. Because now we're just not talking about a toy you mess around with on Saturday, we're talking about something people can actually build products on top of without feeling like the legal floor is moving under them every two weeks. And if you're in this world with open claw, local agents, private workflows, Raspberry Pi, Mac Minis, whatever, Android devices, all this stuff, right? All this stuff, you should be paying attention to this type of stuff. So, let's break it down in plain English for people that might not be super technical. The family itself is

1:32actually pretty smart. Google didn't just make one giant model and call it a day. They shipped four sizes. So, E2B, E4B, 26B, A4B, and 31B. And that is just the parameter size. So, I know the naming looks like something out of Star Trek, but here we go. Once you understand it, it's actually pretty useful. So, E2B and E4B are the edge models. They're the small ones, but not small in like a weak sense. Small in the I can actually run this hardware on my own sense. Like Google says E2B is a 2.3 billion effective parameters, 5.1 billion if you count the embeddings. E4B is the 4 and 1/2 billion effective and 8

2:17billion with embeddings. So, when they say effective, here's the simple version. They found a way to keep the real working footprint lower while still keeping a bigger embedding setup around the model. That matters because it means you could still get more intelligence than the raw number makes you think. Then you have the 26B and the A4B models, which is the sneaky interesting one to me. Total parameters around 25 billion, but only 3.8 billion active during the inference. That means it is a mixture of experts models that does not wake up the whole brain every single time. So, that's a big deal because if you've ever tried to run local models, you know the pain. Everybody loves saying, "Oh, yeah, run it locally." until the fan starts going crazy and

3:02your house catches. The RAM disappears and your machine starts feeling once like it wants to file a complaint. I mean, you need some heavy-duty hardware to get a lot of these local models running efficient and doing actual things for you. The whole point of the 26B A4B is you get way bigger of a model with what with much smaller active slice while it's still working. So, in builder terms, that means better quality than a tiny local model without the same cost profile as running a giant dense beast every single turn. Then the 31B is a dense flagship. That's the one Google is really flexing on for raw quality. About 30 billion parameters, bigger contexts,

3:47better benchmarks, more horsepower. This is the one for the actual work stations, like Mac Studios, Nvidia Spark, or you have 1050 90s running in your house. This is the one for beefier setups. And so, you know, already just at the model menu level, this is useful. Why? Because a real builder stack is not just one model. That's a fantasy. A real stack has layers, different jobs, different cost profiles, different trust levels, different latency requirements. You don't just use the same employee to do payroll, answer the phone, design the logo, and unload the truck. You shouldn't anyways, but it's the same thing that applies here. Real quick, before we keep going, if you're watching this and you want to actually build with some of these tools, not just watch

4:32videos about them, then you're going to want to check out our community down below, Shipping School. We have a full Cloud Code course, a full Open Cloud course, and four live bootcamps every single week, where we actually help you get set up from scratch. Like actually set this thing up, not just tutorial and figure it out by yourself. And we also provide one-on-one coaching, so you could book a call with me, we could share screens, and I can help you get Cloud Code or Open Cloud running on your machine. That's it, no fluff. I built this community because watching YouTube's only gets you so far. We launched it just 3 days ago, and we have over 55 members. You need people around you who are actually building, people who hold you accountable, and coaches who can help you when you get stuck.

5:17I'll put the link in the description down below. Get in now before the price goes up. And Gemma 4 feels like Google finally built a family with that in mind. Let's talk about context window, because this is one of the first things that made me raise my eyebrows. The small models get 128,000 token context window. The bigger ones get 256,000. That's not a fake tiny local context anymore. That means you can actually pass in real documentation, real transcripts, repo chunks, databases, real long instructions, even a whole workflow or a pile of notes without everything falling apart instantly. Google says Gemini 4 uses this hybrid attention setup, sliding window attention makes with global attention plus memory tricks like unified keys and

6:03values and proportional R O P E. Now, I know that sounds nerdy, but the builder takeaway is simple. They are trying to squeeze long context into a hardware normal people can actually touch. That's the whole point of this. And that changes how useful a local model can be because the annoying part with a lot of local models is not just the raw intelligence, it's the feeling that your memory budget is super tight. Like you're trying to run a business with a sticky notepad instead of an actual notebook. Gemini 4 is clearly pushing past that. Then you get into a multimodal support and this part is more important than I think a lot of people realize. All the Gemini 4 models take text and image input. The smaller E2B

6:48and the E4B also support audio natively. And the model card says video can be processed as frames. So, I now you're not just talking about a local chatbox. I'm not. You're talking about something that can actually look at screenshots, parse docs, reason over interfaces, do OCR style tasks, understanding charts and on the smaller models even take audio in. That's the real builder territory because what are a lot of agent workflows really? I mean, when you break it all down. They're not just typing back and forth. They are reading screens, they're extracting structured information, they're listening to short pieces of audio, turning messy inputs into clean outputs, routing actions, summarizing, tagging, translating, and generating JSON. And that is where

7:33Gemini 4 starts getting more interesting for people building actual systems, not just chatting like a chatbot would. Google is explicitly pitching this for agentic workflows. Their launch post talks about native function calling, structured JSON output, and native system instruction. The model card also calls out native function calling support and system role support. That matters because that is the plumbing. Everybody gets excited about the pretty answer. But, if you're building tools, the plumbing is the product. Can the model follow a system instruction cleanly? Can it call a function without getting weird and breaking? Can it output structured data without acting cute? Can it handle an actual workflow instead of just one-off chats? That is

8:18the stuff that determines whether your agent becomes actually useful or becomes a full-time babysitting job. I think we've all been there before. To me, that is why Gemini 4 really matters. It's not just pretending to only be a chatbot. Google is saying this thing is for multi-setup planning, tool use, offline code generation, and autonomous action. Now, do I trust every launch claim from every big company? No, not at all. But, I do trust patterns. And when a company starts optimizing for tool use, structured output, long context, multimodal input, and on-device deployment, all at once, they're telling you what they think what the next wave is. Not a chatbot, not a one-prompt workflow, not just cloud or edge, too.

9:03That's the shift. Now, let's get into the benchmarks a little bit because I know some of you care about that. Most of you do. That's what gets the likes, right? And I think numbers are worth paying attention to. So, according to the official model card, the 381B hits 85.2 on the MMLU Pro, 89.2 on AIME 2026 without tools. It scored 80 on LiveCodeBench version 6. And on 2150 Cord versus Elo. That's not a toy. The 26B A4B is right behind it on a lot of these, 82.6 on the MMLU Pro, 88.3 on the AIME, 77.1 on Live Code Bench, and 1780

9:49on Cord Forces Elo. And even the smaller models are not embarrassing. EB4 hits 52 on Code Bench, E2B hits 44, not terrible. That's saying you fire every cloud model tomorrow and move on to, you know, using these and never going outside. You know, that's not at all what I'm saying. What I'm saying is the floor for useful local intelligence just moved up quite a bit. Google is also bragging that the 31B is the number three option model on Arena's AI's text leaderboard, and the 26B is number six. And that Gemma 4 outcompetes models 20 times its size in that arena. In short, you know, leaderboard talk can get really crazy real quick. We all know that. But when the benchmark story, the

10:34deployment story, and the product story are starting to line up, that's why I made this video. I'm starting to pay attention. And here, you know, they kind of do. Everything lines up. Because the product story is not abstract, Google also launched actual edge tooling around this thing. They have a Google AI Edge Gallery with agent skills, where they're showing Gemma 4 doing multi-step autonomous workflows directly on a device. Querying outside knowledge, making flash cards, creating graphs, connecting to other models even. Building end-to-end workflows through conversation, that is not just a PDF and a leaderboard. That is Google saying, "Here is the model. Here's the runtime path, and this is it for doing useful work with it on actual devices." They also highlighted Light RTLM,

11:20which now supports Gemma 4 across phones, desktops, Raspberry Pi, and more. Plus a CLI with a tool calling support. That's pretty huge for builders, because one of the biggest problems in local AI has never been just the model. It is the last mile, the packaging, the runtime, the inference path, the part where the cool demo is supposed to stop being a cool demo, and start being something actually a real person can use. That is usually where things get ugly. Google is clearly trying to remove some of that ugliness. They even called out Raspberry Pi 5 performance for E2B through light RM with 1,300 tokens per second prefill and 7.6 tokens per second decode. Now, if you're a normal person, you you know, you're not those numbers don't mean

12:05much. And when you're looking at them on a graph, they make much more sense rather than me speaking them to you. But if you're a builder, here's what this means. Google wants this stuff outside the data set. They want it on a desk, on a phone, on an edge box, inside the product. That is where this gets more exciting for us, because I keep thinking about Open Claw. And if you're in our world of the Open Claw ecosystem, which you probably are, Open Claw is the orchestration layer. It is the operating system. It's the routing, the sessions, the tools, the memory, the delivery, the cron stuff, all of that. But the brain inside the stack does not have to be one thing forever. And that is a huge mindset shift for builders. If you're struggling to keep up with content, what I'm about to save you about 40 days worth of work. [music] I built something called Content Machine. It's 10 AI

12:51agents that run on the Open Claw orchestration, and they handle everything. Scripts, [music] thumbnails, X posts, blogs, outreach, clips, newsletters, all of it. So, I went from 1,000 subscribers to 4,000 subscribers on YouTube in 7 days using this exact system. Every single morning I wake up, and the content's already done. I spend maybe 15, 20 minutes reviewing and approving them, and I move on with my day. It works for any niche, fitness, [music] finance, real estate, marketing, whatever you are building, and it is 100% completely customizable to your use case. So, you get the mission control dashboard, all of the cron jobs, everything I've built over the last 40 days helping me gain more and more

13:37people to subscribe and join the community. So, you plug in your own thing and it molds it to you, it learns how you talk and it writes so it doesn't sound like AI slop. $97 one time, it's not a subscription, I'll put the link down below and you'll thank me later. A lot of people still thinking single model terms, pick one best model and stick to it, use it for everything and hope for the best and I think that era is completely dying. With Mythos coming out, Capybara, uh Spud, supposedly OpenAI model coming out, I think things are about to change real fast. So, basically what Gemma 4 gives you is another serious option for like your local brain lab, especially for workflows where privacy actually matters, cost matters, latency matters, or your internet dependency is just

14:22annoying. So, if you think about it like this, maybe your a heavyweight cloud model, you know, still handles your hardest reason, which I do for me. That's how I roll. Your final writing pass or the high stakes decisions, cool, leave it at that. But, Gemma 4 could handle the local prep work, parsing documents, reading screenshots, tagging inbound data, converting the messy stuff into JSON, running a cheap local coding assistant, handling like short audio clips, doing first pass classification, watching a queue, summarizing context before escalation. That kind of split stack is very real now. And it is way more interesting than the usual dumb debate of local versus cloud. That's not even a conversation at this That debate is like arguing whether a restaurant should only use ovens or only use

15:07knives. That's not how the kitchen works, but whatever. Use the right tool for the right state. That's how builders should think now. And for Open Class specifically, I think Gemma 4 matters for these three big reasons. The first one is local fallback, basically the brain is getting good enough to be useful instead of just a cute toy. That matters when APIs fail, when costs spike, when you want more control, when you don't want every single background job touching the cloud. The second thing is multimodal local agents are way more practical now, not just text, screenshots, images, small audio tasks, visual workflows, and that's big if you're building assistance that touch real business operations. Because the real world is messy in general, people send screenshots, voice notes, PDF forms, charge receipts, whatever, all

15:53this stuff. The third thing is Apache 2.0. I was going to redirect when I first mentioned it. Apache 2.0 opens the door for people who want to actually ship a business. And I think this is super the the most underrated part is like a lot of builders don't want vague maybe license, right? They want to know they can build, fine-tune stuff, self-host, package, and sell without having legal stress rash every time the product starts. Google finally moved Gemma 4 into a much more builder-friendly place there. Now, they let me say what I do not think this means. I do not think Gemma 4 means cloud models are dead, not even close. I do not think it means every agent should suddenly run fully local. This is mostly fake macho blah AI slop talk, right? From people don't who don't have to

16:38maintain anything and then are not actually running a real system. And I do not think benchmark wins automatically turn into production wins. We have seen that movie before, right? Where it scores good on the benchmark scores, but it's terrible in actual production use. Function calling support is cool. It does not magically make your tools clean. Structured JSON is also pretty cool. It does not magically make your scheme smart. Long context is cool as well, but it does not magically make your prompts. Open weights are cool, and I love it. It does not magically make your workflow safe. You still need architecture. You still need guardrails. You still need routing. You still need a memory strategy. You still need some taste. That part goes a long way. But the reason I like this release is because it gives builders more room to design better system. So if I was

17:23building this with this right now, here's where my head would go first. I would look at Gemma 4E2B and E4B for edge side assistance, things like mobile utilities and private intake flows, stuff that needs to run close to the user, maybe even offline at times. I would look at the 26B A4B model for serious workstation use, local coding help agent pipelines and faster reasoning where I want more quality without always paying for dense model costs. And I would look at the 31B model when I want the best local Gemma quality possible and I have the box to support it. And if I was wiring this into open claw, I would probably not start by replacing the whole brain. I would just start by giving Gemma 4 a job and start

18:09one job. Then slowly scale it up. Maybe document parser, maybe screenshot explainer, maybe local classifier, maybe just like a simple little code assistant. Um, and that would help keep the workflow moving without hitting a premium API call every 5 minutes. That's how you adopt something like this without turning your own stack into a science project, just a little at a time. To me, that is why this release matters. It gives builders more optionality, more control, more ways to think about their workflow. Not just bigger numbers. Here's the real bottom line. Open models are getting more serious. Edge models are getting more serious. Agent workflows are moving closer to the device and the gap between what was only possible in the cloud and what you can do now with your own hardware is just getting so small. Like

18:54soon we're not even going to know the difference. That changes the game plan for a lot of people, especially people building businesses with AI. Because every time the local layer gets better, you get a little more margin, a little more privacy, a little more control, and a little less dependency on someone else's pricing. That's a huge deal. So no, I don't think Gemma 4 is just another AI news headline. I think Google actually dropped something that builders can actually use as part of a serious stack workflow. And I think the smartest people are going to argue about that all day, whether it beats every other model on Earth. I don't care about that. They're going to ask better question, where does this fit in my work, you know, where does this save me money? Where does this let me keep my data local? Where does this make my agent stack faster, cheaper, more reliable? That's the question that actually matters. And if you start asking that question, Gemma 4 gets very interesting.

19:40So if you want help thinking through this kind of stack, or you want to build with people who are actually shipping this stuff every single week, we just hit 194 members in the last 19 days in the community. It's called Shipping School. It's linked down below. We have eight boot camp calls a week, and my vision for the group is to have 20 calls a week, open claw calls, local model calls, AI agent agentic workflow calls, real estate calls, finance calls, all the calls on how we can implement AI agents to help buy back our time and ask money. So if you're interested in building around other people who are building other workflows, check it out. And if you haven't subscribed already, I don't know what you're thinking. We release four to six videos every single day breaking AI news and open claw use cases and cloud coast stuff. So if you

20:25like that stuff, make sure you subscribe. We'll see in the next one. Peace love.