Dr. Károly Zsolnai-Fehér walks through Meta's Llama 4 models, which claim a 10 million token context window. Scout and Maverick run free on a single GPU, while the larger Behemoth is still training. The models excel at long-context retrieval and handling large codebases, though independent testing is sparse.
Meta's new Llama 4ai just dropped during the weekend and we'll see if it can survive a bunch of challenges. First, it goes without saying that the usual coding vibe check with the bouncing ball works well. Everyone can do the. Wait a minute. The handling of collisions was not exactly correct there. So is this just a one off cherry picked result? Let's check again. Hmm, there it goes again. Okay, so what is going on here? Well, here are two new AI models you can talk to. Scout and Maverick, available right now for free. And the bigger brother Behemoth is still training. He still needs some more reps in the gym. And it is also used to teach the smaller networks and they have more surprises up the sleeve. When we give a bunch of data to the new Deep Seq, it recalls it nearly perfectly all green. Imagine giving it a huge textbook and and then asking questions and when doing the same with Lemma 4. Would you look at that. Is this worse? What happened? Dear fellow scholars, this is two minute papers with Dr. Zsolnai. Well, this happened. Oh my goodness. This has a context length of 10 million tokens. That is nearly 80 times more data given to it than what Deep Sea can handle. That is an insane amount of data. You just chuck in about 10 hours of video and can ask questions about it. Wow. This is a unique property that none of the other AI systems have yet. For text, this feels like infinity. You can talk to it for years and years and it will know your preferences and history really well. That is super useful. And yes, it is not perfect, so it might occasionally forget about your wedding anniversary. That sounds like human intelligence to me. Scout and Maverick fit on a single graphics card, although it needs to be a beefy one or I just rent one at Lambda for cheap and run it there privately. With the long context window you can give it a huge code base and ask for changes. And even though it is not the best at coding, there might be codebases where this is the only tool that can do anything at all. It looks wonderful on benchmarks too, but we don't take those too seriously around here. It uses a mixture of experts model, so it is essentially a committee of little specialized AIs. And that also means that you don't need to have the whole neural network active at the same time. So if you have a high end MacBook Pro or Mac Studio with a little quantization, it runs really fast. The details are in the video description. So that's the good, now the bad. There are already smaller independent studies that are stress testing the context memory and say not so fast. Also, it is not under an MIT license so make sure to have a look at that. So what is it good for? Well this is a free tool for big context projects for many other things, whatever kind of quality and cost pair you desire. Gemini appears to be dominating the Pareto frontier, but for Lemma 4 there is genuine innovation here. Loving the almost infinite text and it also shows how spoiled we are with these cheap AI models. And great news, the future of AI is likely. Models that are free and open. Open science. Loving it. So what would you fellow scholars use this for? Let me know in the comments below. Here you see me running the full deep seq 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. Lamda 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.
