Unsloth Studio + Qwen 3.8-27B on my mini PC

Unsloth, the team whose compressed model files half the local AI world runs on, just released their own app. It runs models, it trains models, it's free and open source. So I installed it live and gave it the biggest thing I could: the brand-new Qwen 3.8-27B, uncompressed, all 55 gigabytes of it.

One command, and a naming correction

The install is a single terminal line. The script detected my AMD ROCm setup, recognized the exact chip (Strix Halo), and pulled architecture-specific builds on its own; a few minutes later the app was serving locally on port 8888. No cloud account involved.

The correction, since I say it wrong for the entire video: the terminal script installs Unsloth Studio, the browser version. Unsloth Desktop is the same interface packaged as a normal installer for Windows, macOS, and Linux. Same UI, two wrappers, pick your poison. Arch and CachyOS users: the Desktop AppImage wants one extra library, sudo pacman -S libayatana-appindicator.

The gotcha: your models don't follow you

I pulled Qwen 3.8-27B through Ollama the night before this video, and Unsloth Studio couldn't see it. Nothing was broken: every local AI app keeps its own model library, in its own format, in its own folder. You either re-download or point the app at a shared model directory in settings. Watch your disk; I now hold roughly 130 GB of the same brain at different compression levels.

That was on purpose, though. The Model Hub lets you multi-select downloads like a shopping list, so I grabbed the whole ladder of Unsloth's Qwen 3.8-27B builds: 2-bit (11 GB), 4-bit (18 GB), 8-bit (29 GB), and the full BF16 at 55 GB. Same model, four sizes. The head-to-head comparison is the next video.

Running the full 55 GB model

Loading it took a few patient minutes and 83 GB of my 128 GB of unified memory, and then it simply worked. Honest numbers from the session:

• Chat ran at roughly 6 tokens per second. Not fast, but this is the fully uncompressed model, something most PCs cannot physically load.
• Deep research mode planned its own steps (you can edit the plan), searched the live web, and delivered a detailed, sourced report in 28 minutes, all locally.
• Vision works out of the box: I dropped a screenshot of my desktop in and it described everything on screen, though it did call Firefox "Chrome Edge". Nobody's perfect.

The app is beta and occasionally shows it, but between the model hub, live system monitor, thinking-effort controls, MCP support, and built-in training, it's the most complete local AI app I've used. I'll live with it before a full verdict.

My rig

Sources

• Unsloth Studio / Desktop docs · downloads
• Unsloth's Qwen 3.8-27B GGUF files
• My earlier letters: the Qwen 3.8-27B release and the local AI Olympics

I am documenting a life of interest and curiosity, tech, money, health, and family included. New video daily.