Full transcript of “Free GPT‑OSS: Local Model with Web Search ≈ GPT‑4‑Mini”. Click a timestamp to open the video at that moment. Need another video? Use the YouTube transcript generator.
0:00 Hi guys, welcome back and today in this video we're talking about all new GPT OSS models which was released yesterday by OpenAI team and we're going to talk about this particular model and how we can use this model within our local machine. So these model as you can see over here has released in two flavors. One is the 120 billion parameters and another one is the 20 billion parameter. These are the two state-of-the-art openw weight language models which is also a reasoning model and it also is available under the Apache 2.0 0 licensing which means it is flexible for you to use it within your local machine and you can use it in an organization based on the licensings and stuff right so you can see that these model that openai claims
0:45 that they are pretty close to the 03 models uh like the GPT3 models and they are quite powerful in terms of reasoning and they also does support the the toolings and stuff. So this is the model details as you can see over here. They also say that the GPD oss 120 billion parameter has got 36 layers of the weights and they have got 117 billion parameters in total and the active parameter per token is 5.1 billion and total experts are 128 and the active expert per token is four right and this is the details that they have got about the performance and steps over here and you can also scroll down a bit and you can see that the the HLE is also given over here which is also quite
1:31 interesting. So you can see that the GPTO OSS 120 billion parameter performs uh with a tooling support of 03 uh like 19 whereas this one is like 24.9 and the GPT OSS um 20 billion parameter is like 17.3 which is pretty close to 04 mini with tools like 17.7. As you can see this is where things are very interesting. So I'm going to show you how we can run this GPT OSS 20 billion parameters uh in my local machine and how it is going to be comparable with the O4 mini which is amazing and you can go to the Olama website over here and you can see that they have got this models and the models the the one this one which is the GP OSS model which is released like 23 hours ago. uh it is one
2:17 of the very fastly downloadable or downloaded version of the model that we have got over here and you can get the 20 billion or 120 billion parameters and I want to show you this one because 120 billion parameter requires quite a lot of uh GPUs to be honest so you can't really run this at least in my machine I can't run this you need at least 80 uh GB of uh of of the GPU memory to run this whereas the 20 billion parameter you can run even with a machine with a RAM of 16 GB which the uh which the OpenAI team climbs. So I'm going to run this one because I have 64GB of RAM. So I can still run this without any hesitations. So let's go and start using this particular model and I have already updated the Olama within my machine. So it's very straightforward for me to
3:02 start using it. So I'm going to jump right into the Olama over here as you can see and I will show you how it actually works. So this is the Olama's new window. If you have been using for some time, this window is going to be quite familiar for you. It's pretty much like the chat GPTs window over here. They have the same kind of the designs over here. Uh but they have some more tweaks this time. You see that there is this option called as uh search. So basically if you're going to be using the GP OSS 20 billion parameter, you also get a search capability which is the tool which is enabled inside this particular feature. Whereas the same thing doesn't work if you're going to go with a deepseek R1 or GMA 3. Those will not have the search capability. Whereas
3:49 the GPT OSS has got the the search as well as the turbo feature. So this turbo feature is very interesting. It is especially for um for you to use an additional power which is powered by the Olama team. So we're not going to talk about that one. It needs additional licensing of $120 per month. But I'm just going to go with the search capability and I'll show you how this actually works. So the moment I ask any questions to the uh GPTOSS 20 billion parameter model. It is going to start doing the thinking, reasoning and tool calling if it requires the search capability. It's going to do the search and then it is going to give you the result pretty much like what chat GPT uh or GPT 40 mini does that exactly the same thing is going to happen even over here within my local machine. So I'm going to say uh what's the weather
4:39 right now in Oakuckland NZ. So the moment I ask this question, see this is not a trained detail or a real-time detail that the model can give you. But because we have got the search capability enabled now this particular model can go and do a search for you online and then look at that. So it's going to do the searching for the current weather in the Oakuckland NZ and it is going to get you that result. This cannot happen if you are using a model like the Deep Seek or Gamma 3 uh over here in this particular um local large language model because it doesn't have the tooling support. But this guy does has the tooling support. So it's going to go and search all this information for you over here. And it's going to give you this information saying as of
5:25 July 20th uh 6:50 a.m. New Zealand time. It's wrong actually because now we are in August, right? So it's giving the it's a July 20th 2025 result. I have no idea why this is giving completely rogue information. So but yeah, you got the idea, right? At least it is doing the search capability, but it's not great search. So you see that it's doing just getting some information there and it's getting that information. But current temperature is 11° in in Oakuckland where the time is 5:30 a.m. So that is what is happening at the moment. Um at least the search is happening. But I'm going to go and do a bit of a coding which I have used with the cloud opus 4.1 to write a playright code and I'm going to use the same exact command over here to see if my local large language model can able to write the code as well. This is something one of the interesting thing that I really wanted
6:11 to try it out and see if the local large language model can compete with an higher level models. I know it's like it's not really an appletoapple comparison but this is like a dinosaur versus an and comparison which I'm doing but I just wanted to see how that really works just for a fun. So this is one of the command that I actually used while working with the cloud opus 4.1 which was released yesterday and I wanted to use the exact same kind of uh informations even in here with the cloud opus 4.1 while I asked to write a full playright C#.NET Net framework using dependency injection page object model separation of concern extended framework run the test in the container for this recorded code from the playright codeen. It actually gave me an amazing code as you can see this is the structure it
6:58 generated and then started creating the project file. It also created the app settings pages uh and all the framework level stuffs over there. dependency injection was also done and then it also started adding a uh docker uh docker container for me for that particular uh question that I asked right the same thing which I'm going to do over here with our local large language model and we'll see how that works so I'm going to go over here I'm going to paste the same command that I actually used for the cloud opus as well and I'm going to hit run and we'll see what is going to happen so this time I expect this particular model to give me a very small information lot not like an amazing
7:43 information like how the cloud opus 4.1 really did but look at that it is doing a lot of thinkings over here and the thinking tag is still running which is also amazing to see that the local large language model is doing all these things like an edge computing is really happening over here without going anywhere in the internet but still it's going to the internet because we have given that search capability there so it's going to do a bit of a search online for the dependency injections and stuff to get some information that it needs. Uh, and also it's checking for the testr runner with the playright c.net on their website. So, it's just reading some information there which is good actually. It's like pretty much like without spending or sparing more um tokens and all those problems that we had before with the chat GPT because you
8:29 also lose the capability of searching after a while if you are using a free version. But now look at that. It's all doing within my local machine and it also found out that we also need to write the playright C# test in the with the docker file. So it is going and searching online uh for the docker setup over here. So it has went to the right link to get the docker information and it has found that over here it's also getting thinking of how we can write it which is awesome to see that everything is happening over here. It's a lot of thinking is going on and I'm just going to wait until this execution is finished to see how it is going to do that and what structure is going to come up with. There we go. I think the thinkings are
9:16 all done right now and now it's creating a project overview. So the test is going to be xUnit. This is the example is fact. And it's creating a fixture for the playright plus dependency injections to create the browsers context page and register page object. and also creating a page object model level code for the login page, application page and the upload page. I think it's pretty much exactly the same thing what the cloud office does yesterday for me. The same thing is happening over here as well. And I think that it's pretty close, right? Like it's doing what it is supposed to and also saying uh why it's layout and things. So which is great. It's creating a configurations with I configuration app settings.json file. I think the same thing goes in here as well with the configurations. I saw
10:01 somewhere about the config. Yeah, look at that. It has created some configurations and it also creates a docker image to run them in the container and this is the structure that it has come up with the CS file app settings program fixture collections. Uh the page it has a base page, login page, application page. Look at that. The same base page and then login page application page. Great. And there is a helpers for the file helper and the locator helper. And then it's writing a test data and the application data. And you know that this new window of the Olama has also got the feature to do a code snippet and stuff. It's also looking pretty great that it's showing everything for us over here. And look at the code guys. Like it is writing all the code pretty much like how it did
10:48 with the with the cloud opus. I cannot really say that it is competing with it. Um again as I told you this is like comparing with an ant with an dinoser but still look at that it is working. My fan is cranking in my laptop already. It is doing quite a lot of things uh behind the scene but look at that the number of tokens it is generating at least I can't see it over here but it's not bad to be honest. It is writing a lot of things for me uh over here and it also adds the key point and the informations for every single operation that it is doing. I think I like this particular code already. This is already pretty good comparing to the models which were came out like in 2023 2024 early 2024. This model is quite good. The reasoning is
11:34 pretty amazing and it is showing me uh the scaffolded code like what I really need for building this test. So I like the fact that this model the GPT oss 20 billion parameter is pretty amazing. it is doing good and they also told that this particular model so if I just go to the browser I don't know why the browser is not here oh yeah it's here uh you can see that they say that this particular openai's GPSS model has got the the features like a genetic capabilities uh which is mainly used for the function calling web browsing uh which is the search that you just saw there and it also does the Python tool calls and structured outputs it also does a full chain of part operation over here and also does the configurable reasoning
12:21 effort uh where you can adjust the reasoning effort like low, medium and high based on the specific use cases and latency need and also uh it has the fine-tunable um details for while you do the customization of the model and then it is of course with the Apache 2.0 licensing but freely without copy life restrictions or patent risk. I don't know what that really means because I don't do any development or things of that nature. There we go. So look at that. This model is looking pretty amazing and I can see that it generates the code and also does a lot of different operation for me uh with just one single command. It's not slow right you you saw that how fast it is. It is performing pretty amazing and I see that this is pretty good. So that's it guys.
13:07 This is all about my take on this new GPT OSS model which runs in my local machine. 20 billion parameter not the 120 billion parameter is never going to work in my laptop unless until I have got a GPU maybe 25090 or 24090 or maybe some other machines but for now I don't have it but but yeah it's working which is which is pretty good. That's it guys. Once again thank you so much for watching this video. Catch you in the next one. [Music]
