I Built a Fully Local AI Video Pipeline (No API Costs)

Okay, so I'm a big fan of the tech channel Fire Ship, of course, as many of you are, too. So, today I wanted to see how close can we get to their videos by using 100% local tools. So, I'm going to walk you through what kind of tools we're going to use for this video. And I want to automate this and let's see how close we can get to his style of videos that I really like.

So uh I decided for automation I wanted to use open code of course and for that we needed an LLM I can run locally and the first thing I looked at was actually trying to use the new Gemma 426B model and I downloaded it. I fired it up but uh yeah it didn't really work as well as I wanted to because the tool calling kind of failed me and yeah it just started running into a loop.

But I did some more research and I looked at this Quen 3.627B model. Uh I think there was this one. Let me see. Uh you can see we have it loaded up here. Quen 3.627B. And this model was very good for this task. So I was super happy with the tool calling. It didn't seem to waste any big tokens uh thinking tokens. So the speed was much better. So this is the model I selected for this

project, the Quen 3.627B. 627B. But of course, that is just the first part that we needed to automate this. And the fire ships mod, the fire ship videos has a lot of like images, memes and stuff like that. So, we needed an image model, right? And I went ahead uh I know I have tried this model before on um on fal said image turbo and I know this is available like open source. So if you go

to said image turbo right we can get this on hugging face too. So I just downloaded this and we can run this locally. That means yeah we don't have to spend any money and you know it's very cheap though but still it's cool to run it locally so we don't need to yeah use the API for this. And of course uh we need a text to voice model for this. And this I kind of already know

what I needed. And this is the hexgrad cooro voice model TTS model 8 82 million parameters. So it's very small and it's super effective on my gx spark here. I'm running this on. So this is super fast too. So that means that we have basically the three things I think we need for this. And the one thing that was missing was of course something we can use to kind of create the the other parts of the

video that I will show you soon. And for that I wanted to try something new that I haven't tried before and that was hyperframes. So this is by hey genen and this writes HTML render video build for agents. Uh this is kind of like the same as reotion. I don't know if you have used that before. I had in my videos but this was kind of the four pillars I wanted to build this on. So we

have right quen 3.76b uh 3.627b 6 27B. We have the said image model, right? We have the Kakoro voice to or text to voice and we have the hyperframe. And of course, we're going to run all of this in open code. So, what I basically did is I wanted to get some kind of head start on the script writing. So, let me show you how I set that up. It's super easy framework that everyone can do.

So, you can't really see it that well here, but basically I looked at a few transcripts from the fire ship's videos to try to understand his humor and the way he kind of structures the video. And I just turned this into like a markdown file. And I passed this along to my open code agent here, right? So we can uh yeah look at kind of some interesting scripts that we like that we want to try to

yeah I don't want to say replicate but bring down bring over the style from right to kind of get this style of videos that uh fireship makes. So now we have also created something called like a markdown file with everything compiled. So if I just do read at and I just call it fake fire local.md and this should have all the information. So let's just do that and then we're going to come up with an ID

uh what type of video we're going to run 100% locally here right uh said image is going to create the images image locally and we're going to compile everything local here without uh spending any money on this. So, of course, this takes a bit of time loading up the model into memory, but when we have that, we can see what we get here. Okay, so you can see now we have loaded the model into memory. It's

not super fast, but it's fine. Maybe it's a bit slow now. I don't know. I felt it was faster this weekend, but let's just let it run. It says read the full file. And now we're just going to come up with an ID for our video. Right. So, the video I wanted to do was to do like a a knowledge here. uh that kind of clawed codecs and all these coding agents are basically slot machines. Uh

I think that's just a funny concept and it's sort of true, right? Because you don't know what you get every time. So basically what I wanted to do is try to create this short video kind of comparing cloud code AI coding agents codecs and stuff to uh slot machines. So I just wanted to see how that turns out in this fire ship uh style we're going to try to to recreate here. So, uh, I think I'm

just going to copy this, um, this post here and we're just going to try to work from there. So, I'm just going to come up with a prompt here. We're going to feed towen and try to work from there. Okay. So, I have the prompt now. So, basically, we are just going to say, uh, we are on Quen 36, not GMA 4. I don't know what happened here. I need to change that. Uh, so we're using

that for script generation. The idea for today's video, we just linked the Reddit URL and we want to compare AI coding agents like Cloex to slot machines, refer to casinos, make some clever jokes, do research, use our surf agent to do some web browsing if we need to. I want to aim for three and a half minute plus and 60% image cards. Good luck. And that is all I'm going to do. So, I'm going to start

this now. I'm probably just going to have to take you back when this is almost done because it's going to take a while. But uh it doesn't really matter. It's local. It can just run in the background anyway, right? And you can kind of see here, right? Uh we're going to do uh script generation, image generation, TTS and render quick start weaknesses. Yeah, you get the point. So I'm just going to fire this off now. And

this is going to take some time, but uh I'm just going to let it run in the background and I'm going to take you back when we are actually done with our first iteration here. Okay, so that was done. That took a lot of time. Uh, I was actually at the gym. So, I let this run in the background, right? So, that is what's so nice about this. And you can see the context window built up

to 174,000 tokens. And we had the to-do list here, but it didn't really complete it, but we have the video. You can see here final we won rendered at AI slots. And if you go to our folder here, you can see we have it here. So, I had a look at it and you can see, yeah, that is rendered, right? So, I'm not going to play it in the Ubuntu here. I'm going to play it like

uh let's play 30 seconds of it so you can kind of see uh the full video here. So, I'm going to play 30 seconds and then I'm going to take you back and talk a bit about it. I think >> last week some guy on Reddit accidentally explained the entire AI coding industry in 600 words and one analogy. Claude Code is a slot machine. That's it. That's the post. And once you see it, you can't unsee

it. It is May 11th, 2026, and you've burned through $40 of tokens this morning. Yes, Anthropic, the safety pill ones, the constitutional a people. They built a digital slot machine, called it a coding agent, and shipped it with a help section. Here is the actual post. R/ better offline top of the subreddit named explicitly for people who think the whole thing is a bubble. You had to use claude code at work. Another AI adoption. >> Okay,

so you can see that was pretty good, right? I'm super happy and it's incredible that this is free now. It's so cool to think about that I can run this like offline on my computer here as long as I have all the sources. It doesn't have to be offline but you get the point, right? I don't depend on any APIs here. And you can see here we kind of ran all the rendering in like hypertext hyperframes

I guess. And if you there's a lot of scrolling here, but uh I wanted to show you. Yeah, here we can see we're generating all the images using uh said image turbo, right? You can see all the the images here. So it took some time, but again I could just let it run in the background. So super happy with this workflow and this is something I'm going to keep uh iterating on and I'll probably make a

version we can run like in let's say cloud code or something too. So more people can try it if you don't have the hardware because the image API is super cheap and you probably have the hardware to run the the TTS model anyway and hyperframes should be no problem. So yeah, that is uh what I wanted to show you today and it's been super fun. I played around uh with this over the weekend. So let me

just show you the channel so you can check out all the other videos. So here you can see we made four videos kind of over the over the weekend. So I'm going to upload the claude code slot machine video. Now I just went ahead to chat GPT and I just created this thumbnail here. Right. So I'm just going to download this just one more prompt and I'm going to upload the video so you can check it

out there too if you want to kind of see the full version. So yeah, uh that is kind of the the workflow I I kind of figured out I wanted to try out this weekend. And again, I'm super happy with it. So, uh one more thing, uh I did create like a new Discord server where I'm just going to be talking about like if people want to talk about AI automation and stuff like that, uh I'm

going to leave a link in the description to this AI automata server I created because uh I get I get a lot of emails, people asking me about stuff. So, I got my Discord back. So, why not just create an server? So, I'm going to try to hang out here and share some of the projects if you want to drop by. So, I'm going to leave a link to that too in the description. And uh just

like a a teaser, I I am working on a project that is I think is even more impressive than this one. Uh but that would be in like an upcoming video. I might show you a thumbnail here. So this is the thumbnail for the type of videos I find an insane way to create that I'm going to share soon. So this is the thumbnail I'm going to be putting up on the video I did create this

weekend. So I think it looks pretty cool, right? So yeah, that's what I wanted to share today. Hope you enjoyed kind of this fully free local AI pipeline workflow that I yeah did over the weekend. And like I said, go check out the Discord if you want to and you can find all the links from the video in the description. So, thank you for tuning in today and I will see you again hopefully very soon and

maybe we talk in the Discord. So, yeah. Bye.