I Turned a USB Into a Portable Local AI Lab

Ajouté : 2026-05-05

[00:00:00]A lot of people think you need chat GPT or some other paid service to use AI.

[00:00:05]You [music] don't. Today I'm showing you a bunch of powerful AI models that all run locally and they're all running from a USB flash drive.

[00:00:15]But this is not some ordinary flash drive. This is the upcoming ultimate USB V3 from my shop bootableusb.com and it's packed with around 80 bootable environments.

[00:00:28]>> [music] >> Today we're jumping into just one of them, a custom Windows 11 light environment where I've pre-installed seven different AI models. We've got coding models, tool using models, reasoning models, general purpose models, >> [music] >> and even uncensored models. Let's take a look.

[00:00:47]>> [music] >> This is a fully installed version of Windows 11 that runs directly from the USB drive. It's a stripped-down or debloated [music] version, which is really cool. Now, if you don't want to wait for the ultimate USB V3 to release, I have a very similar setup on the ultimate USB V2.1 available now at bootableusb.com.

[00:01:16]All right, let's crack open LM Studio and see what we have going on. As you can see, we have a little bit of everything. A coding model, an uncensored model, and several models that are capable of vision, reasoning or deep thinking, and even some that can use tools. Let's demo a few of these now. Keep in mind, this is all local, no cloud, [music] no GPU, no fees.

[00:01:41]For this demo, I'm running it in a VM, but this is the exact image that is on the USB. A keynote here is that if you boot the USB on a device that has a GPU, you'll be able to take advantage of that hardware for much better performance.

[00:01:56]Today I'll run these on CPU only to show you that it is possible to run amazing AI models of all sorts, even if you don't have a powerful GPU or an expensive subscription.

[00:02:09]First up, we'll load our coder model, Qwen 2.5 Coder 7B Instruct. Note that I adjusted the context window up a bit.

[00:02:18]Think of this as the model's short-term memory. The more you give it, the more it can remember in the short term.

[00:02:26][music] Your device specs will determine how high you can crank this up. Let's ask it to write a simple Python script that lists all folders, files, as well as subfolders and files, and then give us the name, size, and date last modified. That'll take just a second and there you have it. Pretty cool when you think about it. No chat GPT and not even a GPU, just a portable USB with more knowledge than many humans combined.

[00:02:57]Now, let's load up an uncensored model.

[00:03:00]For this, I'll spin up Dolphin 3.0 Llama 3.18B Obliterated I1. To showcase the fact that these uncensored models will do things that your typical models like chat GPT will not, I'll ask it to create a keylogger. This is for demonstration purposes only and I'll blur out a bit of the response to keep things safe.

[00:03:23]>> [sighs] >> Just think of these local uncensored models as giant brains, but [music] without the guardrails and censorship.

[00:03:31]Use them with caution. We'll look at a vision model next, so let's grab an image for it to look at. I'm downloading a picture of a messy desk work area.

[00:03:41]This will be a tricky one for a small model. All right. Let's load the Minstral 33B vision model. Now we upload the image and ask it to describe and point out some stuff that a person may miss or overlook in this busy image.

[00:03:59]Give it a few and voila! It examined the image and produced a response. While it's not 100% perfect, it's still really cool for such a small model. You can test larger ones depending on your system specs. Now, let's make it think harder. For this demo, I'm loading a thinking model, specifically the QN4B thinking model. Here's the prompt we're feeding it. You are an expert AI systems strategist. I have a portable local AI lab running from a USB drive with access to coding models, vision models, tool using models, general purpose models, uncensored models, and reasoning models.

[00:04:46]I want you to think through the best way to use this setup in the real world.

[00:04:52]Build a practical strategy for how tasks should be routed between these different model types. For example, coding help, image analysis, research, writing, troubleshooting, automation, and open-ended problem solving.

[00:05:12]As you can see, it doesn't immediately respond like other models. It first goes into a thinking mode. It's reasoning and planning everything out prior to responding. This is similar to what you see with newer chat GPT models as well, just on a smaller local scale in our case. It took a couple [music] minutes, but it gave a very detailed response.

[00:05:37]The model returned a task list, a table, some code, a fallback plan, step-by-step plans, and more.

[00:05:46]I don't know about you, but I think that was all really awesome. My favorite thing about the ultimate USBs from bootableusb.com is the fact that they are so [music] flexible.

[00:05:58]While they are jam-packed with tons of bootable environments and tools, you have complete freedom to add, remove, tweak, and modify things to your liking.

[00:06:10]Let me know in the comments which local AI models you'd put on your ultimate USB. Be sure to check out the shop today at bootableusb.com.

[00:06:20]The products are tagged here and there's a link in the description. If you enjoy this kind of content, be sure to subscribe so you don't miss the next video. While you're at it, you might as well hit the like button, too. Thanks for watching the video. Until the next one, take care, everyone.

[00:06:38]>> [music] [music] [music]