Minisforum AI X1 Pro: Gerbil – Run Large Language Models Locally

In Operation

Launching the AppImage files produces the splash screen below which succinctly summarises the program’s capabilities.

You’ll see that the program supports hardware acceleration with CPU, ROCm, Vulkan and CLBlast backends. As I’m evaluating the software with the Minisforum AI X1 Pro which has an integrated Radeon 890M, I wanted to use the ROCm backend. That’s the AMD software stack for graphics processing unit (GPU).

Try as I might, I’ve been unable to get Gerbil to work with ROCm. Instead, I fell back to the Vulkan backend. There probably isn’t any significant speed disadvantage with using Vulkan (it still does the processing with the GPU).

In the image below, I’m therefore downloading the standard build (rather than the ROCm version).

Once downloaded, you’ll see a Configuration page. I’ll leave the Backend as Vulkan.

After clicking the Launch button, Gerbil proceeds to download the large language model (gemma-3-4b-it.Q8_0.gguf) which is a 3.9GB download.

Once downloaded, I’m then promoted to choose a UI. I went with the Classic UI but it’s easy to change that.

In the image below, I’ve started a conversation with KoboldAI Lite.

You can see that the GPU is being used in the image below.

I set the BIOS to allocate 8GB of system RAM to the GPU.

I’m going to spend some more time investigating whether I can coax Gerbil to use the ROCm backend. If you’ve been successful with an integrated Radeon 890M, please share in the comments.

Website: github.com/lone-cloud/gerbil
Support:
Developer: lone-cloud
License: GNU Affero General Public License v3.0

Gerbil is written in TypeScript. Learn TypeScript with our recommended free books and free tutorials. It’s cross-platform software running under Linux, macOS, and Windows.

Page 1 – Introduction
Page 2 – Running Gerbil

Complete list of articles in this series: