Google Gemma 4: нова ера малих, але потужних open-source моделей для edge-обчислень

Gemma is here. Huge props to Google for continuing to push the frontier of opensource open weights models. I am so happy to say that because not every company is doing that and not every company is doing it as consistently as Google is. So now we have the new version of their Gemma family of models and it's really good. Let me tell you about it. And this video is brought to you by Recraft. More on them later. So, Gemma 4, our most intelligent open models to date. Purpose-built for advanced reasoning and agentic workflows. Of course, when you hear that, what do you think? Yes, OpenClaw. Gemma delivers an unprecedented level of intelligence per parameter. So, these are not massive models. These are actually relatively small models, perfect models to fit on your GPU. and they were able to get such incredible performance out of a small model. And I've been saying this for a while, open- source models are getting smaller, they're getting better, they're getting faster. And that is why I am such a big proponent and so bullish on edge compute on this hybrid between using fully hosted frontier models for the hardest task but for the vast majority of tasks we could probably use compute on any device that we have on our desk. All right, so let me show you the performance. This is just the ELO score. This is not the benchmarks. I'll get to that in a moment. So these are total model size and billions of parameters on the xaxis. Then on the yaxis we have ELO score. What you are looking for is as high up and as to the left as you can go. And as you can see Gemma nailed it. So we have Gemma 31B dense thinking and we have Gemma 26 billion with four active billion parameters thinking. That's a mixture of experts model. Both of which score very very high on the ELO similar to Quen 3.5. But Quen 3.5 that is the 397 billion parameter active 17 billion parameter model. That is a massive model. I can run that but I also have a GB300 which basically nobody has. And so most people can't run that Quen 3.5 model. And now you have a model that is as good and a fraction of the size. You can run this 31 billion parameter model locally. You can do it on most medium to high-end normal consumer hardware. So here's GLM5 coming in way up at the top. Kimmy K 2.5, which I can't even run. I cannot get this even on the GB300 with nearly 750 GB of unified memory. I cannot run Kimmy K 2.5. But within a trillion parameter model, we have right there a 31 billion parameter model performing incredibly well. And then down here we have DeepSeek V3.2 thinking, which is a massive model, but nowhere near it. By the way, where are you at DeepSeek? Come on, come out with your model already. We're all waiting. And also, look at this. Way down here. GPTO OSS, come on. Open AI, it's your turn next. So we have four different sizes of Gemma 4. We have an effective 2 billion parameter model, an effective 4 billion parameter model, a 26 billion mixture of experts, and a 31 billion dense model. Now, I've not actually heard of effective. I didn't know what that term meant, so I actually had to look it up. The E and E2B stands for effective. Smaller models incorporate per layer embeddings to maximize parameter efficiency in ondevice deployments. Rather than adding more layers or parameters to the model, PLE gives each decoder layer its own small embedding for every token. These embedding tables are large but are only used for quick lookups which is why the effective parameter count is much smaller. Okay, so there that is. The entire family moves beyond simple chat to handle complex logic and agentic workflows. Our larger models deliver state-of-the-art performance for their sizes with 31B currently ranking as the number three open model in the world on the industry standard Arena AI text leaderboard. Let's take a look at that. So GLM5, which is a massive model, Kimmy K2.5, which is a massive model. And there it is. Gemma 4 31B, a small model, nearly as good. And by the way, if you're going to use Gemma for commercial purposes, maybe you're building a business that could also use incredible image generation. So, I'm excited to tell you about the sponsor of today's video, Recraft. So, I tested these same prompts across multiple image generation models, and Recraft V4 stood out immediately. And what impressed me the most wasn't the realism, it was the taste, it was the quality, and the control. Recraft V4 understands both short and long prompts. So you can give it a quick one or you can really explain in detail exactly what you're looking for and you're going to get it. So complex compositions with lighting, specific poses, clean typography, even text generation in multiple languages, all done easily. And if you're working on your own branding or UI concepts, the results actually look like a finished product, not a prototype. And with Recraft, you get two separate model families. Recraft V4 for photoreal visuals and then Recraft V4 Vector for fully scalable SVG graphics. So if you're serious about design, you're serious about AI and workflows, Recraft V4 is fantastic. I highly recommend it and you can use it through ReCraft Studio. I'm going to drop the link down below. They've been a fantastic partner. So go check them out. It helps us and it's a fantastic product. Now back to the video. All right, so here are some other things about Gemma 4. Advanced reasoning. It is capable of multi-step planning, deep logic. Great. We already knew that. Improvements in math and instruction following. Agentic workflows, native support for function calling. Yes, this is a model that you plug into your agent. Yes, OpenClaw, you know, I will be testing it. Structured JSON output and native system instructions enables you to build autonomous agents that can interact with different tools and APIs and execute workflows reliably. Code generation. Gemma supports highquality offline code turning your workstation into a local first AI code assistant. Now look, let's be honest. If you are doing coding, you are most likely using a hosted frontier model. If I am writing code, I want to use the best model on the planet and that is either GPT54 or my preference, Opus 46. But you can still do some coding with local models, but it just doesn't make sense to me to do that. Also, all models natively process video and images supporting variable resolutions and excelling at visual tasks like OCR and chart understanding. And the effective models E2B and E4B feature native audio input for speech recognition and understanding. These are tiny models. These are models designed to go on this. Now where it does fall a little bit short and I was a little bit disappointed to see this is context window. The Edge models feature 128K context. That's fine. Okay. So if you're thinking about doing those tiny models, 128K is expected. But for the larger one, only 256K. I really wanted to see more than that. And then they also continue right here. E2B and E4B versions are meant for mobile devices. They have an effective 2 billion and 4 billion parameter footprint during inference to preserve RAM and battery life. In close collaboration with our Google Pixel team and mobile hardware leaders like Qualcomm Technologies and MediaTek, these multimodal models run completely offline with near zero latency across edge devices like phones, Raspberry Pi, Nvidia, Jetson, Orurin, Nano. These are meant to run locally and I think this is just a prediction. We might start seeing them in Apple devices too. So you can find these models everywhere. You can download it in HuggingFace, VLM, Llama CPP, MLX, O Lama, Nvidia Nims. You have LM Studio, Unsloth, a bunch of different options. Download it today. Start playing with it. Start fine-tuning it. You know I am. And let me know what you think. Gemma 4 is released under the commercially permissive Apache 2.0 license. So, go use it. All right. So, some benchmarks. Arena AI text 1452. We saw that. Mmlu, which is multilingual 85.2. 2. Amy 2026 89%, remember Amy 2026 on the frontier is nearing 100%. Life Code Bench 80%, T2 Bench 86%, GPQA Diamond 84.3%. And Steve Vibe on Twitter ran Tool Call 15 across all four Gemma 4 models. And here is what it looks like. So here we go. And there we go. So what you can see here is Gemma 4 31B scores perfect tool calling benchmarks. Really impressive. It is a very good model especially for its size. Go download it. Open source open weights. Go have fun. Let me know what you do with it. And again special thanks to Recraft for sponsoring this video. I'm going to drop links down in the description below. If you enjoyed this video, please consider giving a like and subscribe. and I'll see you in the next