Hello everyone, this is professor

patterns and in this video we're

covering the new GPT OSS model. Now two

models have been released. One is the

GPT OSS 20 bill parameter and we have a

120 bill parameter model. Um both of

these are Apache license so that's

great. U I didn't think it would be

Apache so I'm pretty excited about that.

Um we can see that there's a bunch of

documentation that's been released about

this and the best part is that they

actually did partner with um all of

these local part third party inference

providers. So u hugging face VLM Olama

it's available through all of these

things um from today. So, I've actually

been using this model and uh why don't

we go ahead get started by first making

sure that we can download this model and

then trying a couple of different

prompts to see how well this model

actually performs. Now, I am using Olama

for this, but honestly, I could use VLM,

I could use Llama CBP. Um, I just use

Olama cuz it's convenient. Um, the only

thing is that if you're going to be

using Olama for the first time or

something or after a really long time,

uh, you might want to install and

download the latest version of Olama or

it's probably not going to work for you.

Uh, once you download the latest

version, then you will be able to pull

the model. So, all you would do is just

open up your command prompt window is

type in Olama pull and then the name of

this model. Now, the best part is that

this should also be available on the

Olama website. So, if you actually go to

ola.com/d download and then just search

for the model here, uh you should be

able to see GPTOSS,

the one that you want is 20 bill

parameter. I mean, unless you have

something that can fit the 120 bill

parameter model. Um, personally, I have

a Nvidia 4090 GPU that has about 24

gigs. So, this model is about 16 gigs.

So, um, you can actually try to run it

on a little bit of a less beefier GPU,

but you can run it on a CPU. It's just

going to be extremely slow to run. Okay,

so all I did is I copied that command

and I went here and I just said pull and

I gave it the name of this model. In

this case, I don't want to do what the

120 bill. I'm just going to do the 20

bill parameter. Now, because I already

have this model, it's just going to say

success.

Okay. So after you download the model

now if I go to open web UI directly over

here I should be able to see this model

and interact with it. So I can say

something like hey and I should be able

to get a response here. This model does

support the reasoning capability. So you

should be able to see that as well. Uh

let's ask it to um give me some

resources for learning Python. Sure why

not? Now, in the documentation, they

said that these models are great for

things like agent use and stuff. So, um

I'm pretty excited to actually try some

of that stuff out as well. Um it said

that benchmarks to measure their

capabilities, encoding, competition,

health, agentic tool use when compared

to these other models. So, that's

something that I'm going to be trying

out like right after this video is um

how well these models actually perform

in that case. But here we can see like

it does give really good results. Um we

have a couple of different formats. So

like we have some online course and muks

that it recommended some interactive

platform some supplemental readings. Um

let's say that we wanted to maybe create

a study plan for my Python exam.

Let's see how long it actually does the

reasoning step for. So it does the

thinking. You can see the thought

process here and super fast inference.

Pin down the details. Um, so exam

formats is asking a couple of follow-up

questions and then quick start plan. So

6 weeks, 5 days a week, one to two hours

per day. And these are all the things

that it recommends. So,

okay, not bad. Like I would say like it

and let's see if any of these links

actually work. I'm guessing not. Uh, if

I export to CSV. Oh, nice. That actually

does work. Perfect. Okay. Um, let's also

now try to use see if this is available

through open router as well. So, I'm

going to go to open router and then here

I'm going to search for models and there

we go. GBT OSS and it's

pretty cheap for the 120 build parameter

model. The claude opus in comparison is

$15 per million tokens. And let's take a

look at the

um Claude Sonnet for model is about $3

per million tokens. So this is

significantly cheaper. I'm going to copy

this and I'm actually going to be trying

it for um local

agentic coding type application which

basically is through client. So, I'm

going to type this model in and 120

build parameter model. And I did try to

use the one with Lama, but it just timed

out here. Um, I'm going to try this

again here. But let's try the same exact

prompt that I had before. So, create a

website for my t-shirt store.

Okay, super fast thinking.

And then creating the index.html.

Wow, that is blazingly fast. That's the

Gemini Pro Flash model that it kind of

reminds me of. Oh my god, that that is

insane. But let's see like how good this

is actually going to be because if the

CSS doc is only like 100 lines. Yeah, I

don't know how this is going to turn out

to be. Let's

start the index.html.

Okay. Awesome. Awesome t-shirt store.

And how much did this actually cost me?

Okay. 0.0083.

Let's see if I can say something like

make the website super nice. Make it

extremely awesome. Um, add lots of

interactions. Want to make this website

look super duper fancy. Think step by

step. Well, tip.

And I'm going to say $1,000. Why not?

Okay. So it goes into this whole

thinking process. Um the inference

through open router is insanely fast.

Wow.

Okay. It added a couple more things here

to HTML, but nothing really too intense.

Let's see how much it actually updates

the CSS here.

And so far I haven't even spent a scent

yet. So that's going to be interesting.

There we go.

I'm guessing that once more people start

using this model. Um because if you take

a look at the providers here, let's

click on this

and who is actually providing this

model. So we have it through Grock

Sirius and Fireworks.

Okay. So I'm guessing the more people

that will watch this video and also like

the more people that will start using

these models, it's going to reduce the

overall latency time here. But for now,

it is running super fast.

Okay. Awesome t-shirt store. It's got

these sort of like cool interactions.

Not sure what this dark mode thing is

about. Um, yeah, not amazing, I would

say. Like I guess I can see why this

model is a little bit on the cheaper

side, but really fast. I'm trying to

think maybe I could pro probably use

this model for something like really

quick checking for bugs or something. Uh

but you know I don't really see me using

this um

as a replacement for something like

Claude Sonnet or Opus or something for

coding tasks. Now there is also like

fine-tuning um stuff that's available

here. So if you actually see there's an

article for fine-tuning with GPTO, OSS

and hugging face transformers

and they actually have a full

documentation here. Um you can set up

install all the different libraries,

prepare the data set and then fine-tune

this specific model for your

application. So I'm also going to create

another follow-up video for fine-tuning

this model. I think that would also be

interesting.

But these are just some preliminary

projects uh that I've been working with

with the GPTOSS model. Try the local

variants. There's also the if you go to

gpt

oss.com.

Um you'll also be able to see this sort

of interface where you can try out like

different things like reasoning levels,

show reasoning, don't show reasoning,

and try these two models out as well.

Give it a couple of prompts like how far

away is the sun from Earth. This way you

don't even need to download this model

locally on your computer. um it will

just kind of run here through this

interface.

But yeah, that's pretty much it for this

video. More to come. I'm going to be

covering this in a lot more detail. I

think this is a really exciting update.

Um it's not I'm not mind blown or

anything like that based on like how

well this model is performing so far.

It's decent. It is fast and I like the

fact that it's local, but does it

actually replace my existing workflow?

Probably not. But let's see. I'm still

going to be experimenting with this a

little bit more and I'll keep you

updated. Thank you all for tuning in.

This is Professor Patterns and have a

good rest of your day. Goodbye.