Gemma 4 on MacBook Pro M1 Max is INSANE! (Local LLM)

What's going on everybody? This is David and today we're going to be checking out Gamma 4 on my M1 Max MacBook Pro 16-in.

Now, I'm somebody that's just getting into local LLMs. I recently downloaded Ollama on this MacBook and this is just

something that I'm exploring just for fun. I did dabble a little bit in one of

fun. I did dabble a little bit in one of the cloud services like Gemini and ChatGPT and you know, Claude and all that. So

far my favorite has been Gemini, but that's just I feel like for like creation work it's pretty good for that.

But I you know, I've I've been just dabbling in local LLMs just because I don't like sharing information with cloud services

and stuff like that. So if I can do you know, like research or just stuff like that locally, I would much prefer that. But I am a beginner, so

that. But I am a beginner, so there's that. I'm not going to know the

there's that. I'm not going to know the ends and outs of what these models do and whatnot. So I am going into this as

and whatnot. So I am going into this as a beginner and just yesterday Gamma 4 has been released which is from Google. So

me being somebody that has been you know, mostly using Gemini this is something that interests me and I do have Ollama like I said, so I'm going to download it today. I'm going to

download the biggest model, the 31 billion parameter one.

I do have 32 gigabytes of unified memory, so I believe it should fit in within my memory pool.

I think there's like 24 gigs that I can use with a model and the things that I'm going to do is I'm going to probably just kind of ask it pretty simple

questions. Maybe I'll ask it you know,

questions. Maybe I'll ask it you know, what's in an what's in a certain image?

Maybe give it some, you know, like a title or description or tags for this video in particular just to see what it spits out. Maybe I'll even use it for

spits out. Maybe I'll even use it for this video, who knows? And I'll just kind of experiment. You know, maybe there's like like a really hard prompt that I can ask

it as well. We'll see. Let's go ahead and check it out and see what we find.

All right, guys. Here we go. So, we have Ollama open up on my MacBook here and Gemma 431 billion parameter model. We do

have the context at 16K just so I don't run out of VRAM, but obviously if you have more VRAM to spare, you can use a a

bigger context model or you can use the 27 a billion parameter and then increase your context up to you whatever you want to do, but what we're going to be doing

here is just running the full fat 31 billion parameter model. And then I'm just going to ask it a few prompts here and then I'm also going to ask Gemini,

the thinking model of Gemini, the same prompts. Just to see, you know, what it

prompts. Just to see, you know, what it spits out. So,

spits out. So, uh the first thing I'm going to do is I'm going to ask it what the uploaded photo is.

And that photo is going to be of a game called Marathon. This photo right here.

called Marathon. This photo right here.

Just going to upload it. And I'm going to ask it um this right here. I have uploaded an image of a game. Describe what is in

this image and of what game it is. Let's

go ahead and do that. And then I'm going to do the

do that. And then I'm going to do the exact same thing for Gemini.

And there we go.

Let's turn this off. Do we do that and ask it the same thing.

And let's wait for its response. So,

I'll be right back after this is finished.

All right, guys. Here we are. So, the

local um Gamma 4 31 billion parameter model has analyzed the photo and it thinks that this game

is Half-Life Alyx, which is hilarious.

Uh the reason I think that before even spat out the There it is.

Uh the answer is because I was looking at its thought process, which was hilarious. This is actually so

hilarious. This is actually so interesting to see how these um how these AIs think. But, basically,

it kept thinking that it was Half-Life Alyx and then it's like second-guessing itself. It's like, "No, the art style is

itself. It's like, "No, the art style is too stylized. Is it Superhot VR? No, the

too stylized. Is it Superhot VR? No, the

environments are usually white. Let's

try Half-Life again. Actually, wait.

Look at the colors. This looks like Half-Life Alyx's gravity gloves." And

then it's like, "No, the the gun isn't a gravity glove." So, it's like

gravity glove." So, it's like second-guessing itself constantly.

And then it's like, "Is this a mod?" So,

anyways, the description it gave us is that the image is a first-person view from a video game. In the foreground,

video game. In the foreground, player is holding a stylized pistol that has been colored in vibrant neon pink and green. Attached to the top of the

and green. Attached to the top of the weapon is a white cube with black QR code-like markings.

And then it has um identified the text 750R.

But, it thinks that this game is the VR game Half-Life Alyx. I know it's not. I

played quite a bit of Half-Life Alyx.

So, it uh has wrongly guessed that this is from Half-Life Alyx, which it is not. Let's go to Gemini. Gemini

Gemini. Gemini tells us that this image showcases a first-person view of a player character holding a distinct, highly stylized weapon

from the upcoming extraction shooter Marathon developed by Bungie.

Now, they say that it's upcoming, but actually the game has already come out. So,

that's kind of wrong.

It's giving us a description of the weapon, what it looks like.

And here it's telling us that based on the file name and the visual style, this image appears to be part of a press kit

shown showcasing pre-order rewards.

And that is correct. This is from a press kit. I downloaded that image from

press kit. I downloaded that image from a press kit. So, that is very interesting that it told us that. But it

did identify identify that this was Marathon.

So, that is one win for Gemini. All

right, the second prompt that we're going to give it is that I have made

Let's delete this.

I have made a video of the new local LLM named Gamma 4 running on my MacBook Pro.

The video will be uploaded to YouTube.

Please let me know the best click-through rate for YouTube titles, description description tags, uh tags, because you have different tags,

community post, and pin comment for this video.

So, that's a lot of information for it to process, right?

So, I'm going to start a new chat here with a Gemini, and I'm going to ask it the same exact question.

And go.

All right, and we'll be right back after this is finished.

All right, Gamma 4 has finished in 70 seconds.

Not too bad. Gemini finished in like 5 seconds, but here we go.

So, this one is giving us option one, option two, option three, and option four for the titles. It's giving us the description like we asked.

It's even giving us a timestamp.

Some links I can link to, which is pretty nice. It's giving us the

pretty nice. It's giving us the description tags, the video tags. It's

telling me the pin comment like I asked it to.

And it's giving me a community post as well. So, that's one of those things

well. So, that's one of those things where uh if I post like you know, um like uh

uh a YouTube post asking people like, do you like this or that? Like a poll or something like that.

So, that's pretty cool.

And then, let's see what a llama is telling us to do.

So, it's still giving us the answer here, but let's just start out at the top.

So, it's saying that note, as of my current knowledge, Google has released Gamma and Gamma 2.

If you are referring to a new release or a specific version you've dubbed you've dubbed gamma 4, the phone strategy will work perfectly.

If it was a typo for gamma 2, these templates still apply.

Okay, so gamma 4 doesn't even know it's gamma 4. That is quite interesting.

gamma 4. That is quite interesting.

It literally does not know what model it is. It still thinks that we're in gamma

is. It still thinks that we're in gamma 2.

And there's actually even a gamma 3.

Okay.

Interesting, interesting. So it gave us option one.

And then it gives us three titles for option one.

And then it gives us a description, okay, with uh timestamps, hardware, description, tags, and tags.

All right, doesn't give us any type of community post or anything like that. It

actually gives us an option two with three additional titles, a whole different description, and tags, and an option three with three different

titles and a whole different description. Well, that's pretty cool.

description. Well, that's pretty cool.

This actually uh gave us more information than Gemini. Gemini just

gave us four op- four different titles, but Ollama or gamma 4 gave us three different titles for

three different types of themes. So that's interesting.

Okay.

Okay, it did give us the community post, which is kind of cool.

All right, so it did do that, and it did give us a pinned comment.

Ah. All right, so this actually gave us pretty much everything I asked it to do and even more.

It literally gave me more than I asked it or more than Gemini did.

So, I would say that's a win for Gamma 4 except that it didn't really know what Gamma 4 is even though that's the model I'm asking, so that's a little

strange.

Anyways, let's go ahead and uh, give it our last prompt.

So, this is a prompt that I asked Gemini to give me to confuse local LLMs and apparently this is supposed to be pretty

hard for local LLMs to you know give a description of, so I'm just going to put it into Gamma 4 and then I'm going to do the same with

Gemini here.

Interesting interesting.

So, let's see what they spit out. This

one is it's so interesting to see how these AIs think, you know what I mean?

Like it's it's almost like peeking into somebody's brain or thoughts, you know?

It's kind of crazy.

Anyways, let's come back to when it gives us the answer.

All right, finally finished. It took it 200 seconds.

Let's see what it says.

This hue resides beyond the reach the human retina.

Kind of saying the same thing that Gemini is.

It is a ghost of a pigment, a transport transparent void occupying the center of the room.

Despite its invisibility, it possesses a radiance that dwarfs every other object.

Its presence is felt as a searing pressure, a brilliance so intense that it washes out the saturation of the surrounding walls.

All right. I don't know if that's right or right or wrong.

Uh Gemini here says it occupies a position on a specific on a spectrum where whatever this means. Placing it out of

beyond reach of the rod Okay, I guess Gemini I was more detailed.

I honestly have no idea uh the answer of this question, but Okay, it did give us different types of answers though.

That's interesting. All right. Well, I'm

just playing around with it. What do you guys think of this new model? Let me

know in the comments below and uh have a blessed day. Peace out.

blessed day. Peace out.