In this Two Minute Papers episode, Dr. Károly Zsolnai‑Fehér reviews Google DeepMind’s newly released Gemma 4 family of large language models. He emphasizes that Gemma 4 is a free, open‑source, densely‑parameterized model that can run on modest hardware – even a phone or a first‑generation Nintendo Switch – thanks to its small memory footprint. The presenter highlights four technical innovations that make the model surprisingly strong: (1) highly curated training data, (2) a hybrid attention mechanism combining sliding‑window and global attention, (3) improved raw‑image handling, and (4) a shared key‑value cache that reduces recomputation. He contrasts Gemma 4’s dense architecture with the more common mixture‑of‑experts (MoE) models, noting that despite being dense it rivals much larger models. Additional advantages include a 256 k token context window, Apache 2.0 licensing that removes usage restrictions, and strong agentic workflow capabilities such as tool use, local coding, and custom instructions when integrated with OpenClaw. The video also mentions practical deployments (offline translation, summarization, real‑time image classification) and community uptake (10 million downloads in the first week). Limitations are acknowledged: no live browsing/database, occasional hallucinations, difficulty with very high‑frequency visual details, and challenges on complex open‑ended tasks. The presenter concludes that Gemma 4 represents a lasting, freely accessible AI gift to the broader public, contrasting it with proprietary, subscription‑based services.
Chapters
Google DeepMind gave an amazing gift to humanity, and it is full of surprises. Here's why. Today, we are living in the age of AI, where these smart assistants and agents can do things we could only dream of 10 years ago. But many of these solutions are proprietary, require a subscription, and run in the cloud. And then, this happens. Yup, some OpenClaw users reported losing access to their Claude AI subscription, citing heavy workloads. Now, maybe they did something unsavory, I don't know. I also understand that you pay a fixed rate, so you can't eat all you want. I respect that. However, this is the point. We have to rely on the goodwill of these companies for our workflows. So this is why I keep saying over and over that we should always look for options where we own these AIs and run them on our own systems for free, forever. No one can take them away. NVIDIA came out with their NeMoTron3 Super, which has super capabilities, but its hardware requirements are also super. Not so much with Google DeepMind's new AI, Gemma 4. This is a free and open family of models, and yes, finally, the smallest ones require only a few gigabytes of memory. No need for an expensive GPU. So much so that I wanted to wait a bit before publishing this video to see how you fellow scholars use it in practice. And— look at that! It runs on your phone without an internet connection. And folks are already using it in practice to create offline translation and summarization apps. Also, real-time image classification running in your browser while talking like a bard? No problem. You can also fine-tune it with maths work. It is so good, it has a little ecosystem around it already in just a few days. Wow! Because of the brilliance of you Fellow Scholars. Nice work! But it gets better. You see, the smallest Gemma is so small it runs on... Oh my! Look! I love that! It runs even on an old, beat-up Nintendo Switch. First generation. Not exactly something with a lot of memory or processing power. Still runs the 2 billion parameter Gemma-4 model. Now that is a gift to humanity. But it gets really strange from here on out. Here are 4 things that I found really surprising. Dear Fellow Scholars, this is Two Minute Papers with Dr. Károly Zsolnai-Fehér. One, they also have a bigger, 31 billion model, which was the third best open model. And now, hold on to your papers Fellow Scholars, because it beat some models that are 10 times larger, and still competitive with some that are 20 times larger. Wow. On some measurements. And it is a dense model. What? What is going on here? You see, many of these modern AI systems you encounter are what they call a mixture of experts model. MOE. These are huge AI models with many parameters, and to make sure we don't burn down all of our hardware using them, it splits up the big brain into many small ones. If you have a biology question, it chops it up into small parts and routes them to the parts of the brain that it thinks are the best experts at processing it. Typically to the top 2 to 8 experts. Ask only them, not the whole brain. Yes, with that we only activate one small part of the brain at a time. Heh, it makes sense, right? It's not a simple process, but It is possible. This enables us to create huge intelligent models that are still efficient. Dense models, however, just light up every parameter of the system. Hm. These are not new. And in some ways, these are very inefficient. You light up all the 31 billion parameters in the brain all the time, no matter how simple or complex the question is. But this one? This one is somehow magically good. How? They did 4 amazing things. One, Google didn't just dump half the internet into it to learn from us. No-no. They apply super strict filters to give it only highly curated training data. Hmm. That is actually really good advice for our thinking too. Don't let everything in. Curate your information diet. There is a lot of noise out there. Ignore it. That is excellent. Two, they use an interesting attention mechanism that has a sliding window and also global attention at the same time. Okay, what does that mean? Well, when you read a book, you read it line by line to finish a page. That is a local sliding window. With that, you get all the details. But sometimes you want to zoom out and ask, okay, what book are we reading? Which chapter is this? That is global attention. Here they use both and call the mechanism hybrid attention. 3. It is better at understanding images. You know, Gemma 3 had weird glasses on, and its image understanding was Kind of a lie. If you gave it a landscape image, it squished it back into a square image before processing it, losing some information. It squishes everything into its own preconceived box. Not good. Gemma 4, however, understands the image as is, and the difference really shows on any benchmark that has to do with images. 4. It has a shared KV cache. KV cache is short-term memory for what you are currently talking with it, Documents, questions. Now, the layers of this neural network like to recompute their fresh memory from scratch. This one doesn't. It essentially borrows the memory already computed by earlier layers. Less work, nearly the same result. Wow. This is one of those ideas where we are wondering why we didn't always do it like that. Okay, and all this was just part of my first surprise. Second surprise: it is fantastic at agentic workflows. This is where we don't just have an AI assistant that spits out a bunch of text. This is when we give it arms and legs and ask it to do stuff. Tool use, local coding, and a ton more. Plug it into OpenClaw, and it can book a plane ticket, look for news, and summarize it in a more unbiased way. Or— write silly emails to Károly from Two Minute Papers. That sort of thing. It is really good at that. So, when any company decides that you can't use their system anymore, that's alright, just plug in Gemma 4 and you're good to go. For free. People find that if you give it custom instructions, sometimes you don't even notice the difference. That is huge! Surprise number 3: the context window was improved to 256k, twice as big as Gemma 3 had. This is pretty expensive to compute, so don't take it for granted. Here you are not going to chuck gigabytes of movies into it, but for a few long documents, it is perfectly fine. 4: The license. Oh my! The license! This one gets overlooked so much. Gemma 3 came with a Gemma license. In other words, it comes with strings attached. The model comes with handcuffs, if you will. If you use it to create training data for a derivative model— yep, that one inherits the handcuffs too. But! With Gemma 4, not anymore. Look, Apache 2.0 license. Now we're talking. Yes! This license is true to the open source spirit. You can modify it, sell it, deploy it commercially with almost zero friction. Make derivative models do a ton of stuff with far fewer restrictions. This is huge. Thank you so much. Now, not even this technique is perfect. For instance, the model does not have a live database. Without an agent harness, it cannot browse or look stuff up. Meaning what? Well, meaning that it can be confidently incorrect. The internet special. Also, for highly complex open-ended tasks, it's not great at that. Or when you have images with lots of high-frequency visual details, thin structures, blades of grass, or a fence from far away. Not great at that. It's going to need even better glasses. But adding this all up, this is an incredible gift to humanity. One that cannot be taken from us. This is not for Mr. Moneybags, this is for the little man. And it is free, for all of us forever. Hugely appreciated. Absolutely loving it. What a time to be alive! Also, I waited with this video because I did not want to just take the marketing messaging and copy-paste it to you. I wanted to see how you Fellow Scholars are actually using it in practice. Read through your experiences with it. Does it really work in practice? Super important. That's what we are here for, not for the copy-pasted media headlines. But that needs time. Trying to explain all this in simple words also takes time. I don't have a team here, I do everything here with the help of my wife, from the writing to recording, video editing, everything. I am trying my best here, but I think it gives you more accurate information, and that is the most important for me. So now, after 10 million downloads in the first week and more thorough testing, my opinion is that yes, this thing works. I would like to send a big thank you to every single scientist who worked on this. And hold on to this one for dear life, because the Frontier model just got locked down for a few select clients. You know, it's a big club. And we ain't in it. If you enjoy this, consider subscribing and hitting that bell. Here you see me running the full Deepseek AI model through Lambda GPU cloud. 671 billion parameters running super fast and super reliably. This is insane. I love it. And I use it on a regular basis. Lambda provides you with powerful NVIDIA GPUs to run your own chatbots and experiments. Seriously, try it out now at lambda.ai. Papers, or click the link in the description.
