Transcription
Looks like the rumors from yesterday were true. Google has introduced the most capable model that they've ever released that can run on a single GPU or TPU, called Gamma 3.
Gamma 3 is a new collection of lightweight, state-of-the-art, open AI models. You heard it: open AI models. These are models that are built on the same technology as Google's Gemini 2.0. It is designed for efficiency and offers four models with varying sizes: 1B, 4B, 12B, and 27 billion parameter models. Which are all optimized for running on various devices, including phones, laptops, workstations. And it's actually pre-trained in 140-plus languages, with native support for 35-plus.
The best part is, guys, these are open-based models, meaning that you can install this locally with Llama or LM Studio. These are not just simple text-based models. Excluding the 1 billion parameter model, the other three models are multimodal infused models; that means that they can support text, images, and short videos. It supports up to 128k tokens, except for the 1B model, which supports 32k. But in regards to its performance, it's quite impressive for its size, as it can outcompete larger parameter-based models like DeepSeek version 3 (that's 671 billion parameters) or Llama 3's 405B. What's also insane is that Gamma 3 also outcompetes LLaMa 03 Mini, which is just insane. When you're looking at this graph over here, it's kind of funny to see that it only requires one single Nvidia H100 GPU, whereas DeepSeek, as well as these other models, require multiple.
In comparison to its previous model, the Gamma 2, you can see that there's a huge difference in almost every benchmark in all of the sizes, where it outcompetes it on almost everything, recording a higher score in math, coding, general Q&A, logical reasoning, and so much more. If you want to deploy it on your phone, web, or cloud, I'll leave a link to AI Google's dodev website, which will help you get started.
Now, to install this, it's quite easy. You can easily access all the endpoints off of Hugging Face, or you can install this using Llama. You can easily install this locally with your operating system, and then you can go over to the model card, select the model size you want, and obviously make sure you have the prerequisites to run it. And once you have figured that out, you can then copy the model card command, and then you can go over to your command prompt. Once you have opened your command prompt, you can simply go ahead and paste it in the command and then click enter, and it will start installing this model locally in LM Studio. It's also pretty easy. You can go ahead and search up the Gamma 3 model, and you're going to be able to find different GGf, uh, versions of it, and you can install the Quan size that you would want by simply clicking on the download button, and then you can access it within their chat interface.
Now, say you want to chat with it right now on the web. You can easily do so with Google's AI Studio, completely for free. You can click on the model selection card, scroll all the way down, and then select all the Gamma 3 models. In this case, I'm going to be using the Gamma 3 27B parameter model. And what we're going to do now is assess it off of these different prompts. We have eight different prompts that will assess this model in different categories: from coding, logical reasoning, and having the ability to assess how well this model is in terms of generating these different types of answers.
Before we get started, I just want to mention that you should definitely go ahead and subscribe to the World of AI newsletter. I'm constantly posting different newsletters on a weekly basis, so this is where you can easily get up-to-date knowledge about what is happening in the AI space. So definitely go ahead and subscribe, as this is completely for free. So let's get started. The first prompt is to have it build out a simple web app with the ability to use HTML, CSS, JavaScript, and essentially this web app is to log monthly expenses and income and have it visualize the data. So this is going to be something that prompts how well the model can generate structured, functional code. It's going to also evaluate understanding of UI designs, data handling, and visualizations. So, it has finally finished generating the code. We can now take a look at it, and this is our personal Finance Tracker app. And I'm kind of surprised to see that it was capable of generating a pretty detailed-looking app where you have the ability to log in your transactions, and then you can add a description, the amount, income, utilities, and then you can simply go ahead and add it. This is where you can then track it with the financial summary, and then you have your transaction history. And all of this was built with one single prompt. I'm definitely really surprised by that, and I will definitely give this a pass. But this is just amazing to see that it was capable of generating that with its parameter size.
Next up, we're going to be assessing this model of its image understanding capabilities, where it's going to be focusing on evaluating multimodal abilities with text as well as image analysis. So essentially, we're going to have it create a short story based off of the couple of images that are attached within this prompt. We're going to see how well this model is in terms of object recognization as well as scene understanding. So we can see over here this is an image that we're providing of a dog with a croissant, and you can see it is working on generating a short story for it, and then over time it will compile all of these different images and create a story based off of all of it. Pretty impressive that it's already doing a great job in generating the story for us with the wolf. You can see it's focusing on that, and it's focusing on the treasure image, and then over time it'll focus on all this. So this is definitely deemed a pass.
Next, next up, we're going to have it generate SVG code to create a butterfly that is symmetrical with wings and a simple styling. So this is actually a prompt that tends to fail with most large language models; only a few have been able to do this. We're going to see if it's able to generate the SVG code. So I've now copied the code, and I'm going to place this into an online SVG viewer, and unfortunately that does not look like a butterfly. I get it; it does the symmetrical part, the antenna's right, but the butterfly is nonexistent in this. So unfortunately, this is actually a fail.
The next prompt is related to mathematics, where it's a simple algebra equation, and we're going to have the model problem solve and solve for x in this particular equation. We're trying to see how old this model is in terms of problem solving, evaluating quadratic equation solving, and essentially the answer is x = 3 or 1, which is 100% correct in this case. So with math, it's definitely quite impressive, not just based off of this test, but something that I saw with other tests that I had done before this video.
The next prompt states that a farmer has 10 cows, five goats, and two chickens. Each cow gives 10L of milk daily, each goat gives 3L, and chickens do not produce any sort of milk. How much milk does the farmer collect in a week? So essentially, this is a logical reasoning prompt. We're trying to see if it can use logic in deduction. You're going to see that if it's able to use multiple steps, which it does, and it is actually providing the correct answer, which is 885L. So overall, it did a great job in its logical deduction. So let's go ahead back to the benchmark test and give this a pass.
Next up, we're focusing on a prompt that assesses how well the model is in terms of debugging and error analysis. So essentially, this is a Python function that is supposed to return the sum of all even numbers in a list, but it has a bug where it doesn't, or it adds an odd number to the mix. And you can see right away that it does correctly fix it by making the equal one amount to zero so that it is actually even, and you can see that it explains this where it was able to find the faulty code, it fixed it, and you can see that this model does a good job. It's obviously a basic function, but it was able to identify and correct the error. So in this case, this is definitely deemed a pass.
Let's now assess if this model has any sort of common sense. This is a common sense reasoning prompt, which is going to focus on general knowledge, and I'm asking it: if you put a bowl of water outside in freezing temperatures, what will happen? Explain why. So trying to see if it has basic physics knowledge as well as evaluating casual reasoning. So you can see it is focusing on the science behind it, which is talking about the temperature and molecular motion, freezing point, heat transfer, etc. It's also talking about the process, cooling, freezing begins after it reaches 0°C, latent heat of fusion, etc. So this is definitely deemed a pass because it was able to explain the science as well as the process of it. So let's give this a pass, and this is deemed a pass.
Now, overall, this is definitely a great model for its size, but obviously from the prompts that you saw, they're not super intricate or super hard, and this is because this is the type of prompts that you would send for a smaller base model. Obviously, it has the ability to compete against something like DeepSeek version 3, which is definitely nice, but it's still not at the level where you can assess it on harder-based prompts. Obviously, the SVG code is a difficult prompt to work with, but overall it's a great small-based model that you can deploy locally, and it's something that can easily be accessible with one single GPU or CPU. As this is a model that's meant for something that's capable to be used for most consumers, it's something that's open, and I definitely think that this is a great all-around model. It's a little lackluster in coding, but in math, multimodal, and general knowledge, it's definitely impressive, and it has a lot of benefits to it in those categories.
If you like this video and would love to support the channel, you can consider donating to my channel through the Super Thanks option below, or you can consider joining our private Discord where you can access multiple subscriptions to different AI tools for free on a monthly basis, plus daily AI news and exclusive content, plus a lot more. I'll leave all the links that I use in today's video in the description below. Make sure you take a look at this because this is definitely a powerful model despite its size. So that's definitely quite impressive. But with that thought, guys, thank you guys so much for watching. I hope you enjoyed today's video. Follow me on the newsletter, follow me, enjoy the private Discord, make sure you take a look at the Twitter page, and lastly, make sure you guys subscribe, turn on the notification bell, like this video, please take a look at our previous videos, because there's a lot of content that you will truly benefit from. But with that thought, guys, have an amazing day, positivity, and I'll see you guys fairly shortly. Peace out, fellas.