One of the more popular locally hosted AI assistants would be OpenClaw. It’s a free and open-source AI agent that can make use of LLMs to generate responses. It also supports both cloud and local AI models, and I was curious to see how well it would fare on a relatively high-end laptop.
To this end, I deployed OpenClaw on my Arch Linux install, intent on finding out its limits (and uses). I hit a wall quickly, but the journey itself was fascinating (and educational!). All things considered, I’d consider OpenClaw a worthy assistant to have on your PC, local or not.
Meet the test laptop
Ryzen AI Max goodness and quite a lot of usable VRAM
The laptop of choice was the ASUS ROG Flow Z13 (2025). My version came with the lower-end Ryzen AI Max 390 and 32GB of unified memory — shared between the system RAM and the iGPU’s VRAM.
This particular setup has its benefits, including the ability to allocate over 16GB of VRAM to the machine. With over 16GB available (24GB maximum if you enable the BIOS setting), you can run some rather large LLMs that would otherwise choke on an 8GB card.
That being said, it’s not perfect. The laptop’s Radeon 8050S iGPU is roughly comparable to an Nvidia RTX 4060 Mobile in real-world horsepower, and that translates to AI performance as well. In fact, the 4060 Mobile edges it out in some cases, but the 8-gigabyte hard VRAM limit will always be a serious bottleneck.
The 8050S also uses ROCm or Vulkan instead of Nvidia’s proprietary CUDA stack, which may or may not be a deal-breaker. That said, it’s still a mighty fine system that can run local AI models reasonably well.
As for the test environment, it was a fairly vanilla Arch Linux installation with KDE Plasma and the requisite AMD GPU drivers. All tests were conducted with the max TDP (93 W) set in place, with fans kicked to maximum speed in order to ensure no thermal shutdowns. CPU boost was also left enabled.
Installing OpenClaw on an Arch Linux system
Surprisingly complicated if you start from scratch
Installing OpenClaw on a fresh installation of Arch Linux was surprisingly complicated. For starters, I had to install Ollama first. Then, after installing the ROCm and Vulcan-specific packages, I enabled the service, and it was good to go — but I still needed to install a local AI model.
I intended to test OpenClaw with a local model first, and so cloud-based options were out of the question. As such, I chose Gemma 4, with the 12b version in place. The reasoning was simple: Gemma 4 is decently usable on local hardware, and the 12b version can take advantage of 16 gigabytes or more of VRAM.
This made it an ideal fit for my AI Max machine, and the unified memory pool helped a lot.
After one installation, I went back into the OpenClaw setup, following the steps on its Arch Wiki page. I opted for a manual setup, selecting Ollama local as my target. A few interactive prompts later, I was booted directly into the OpenClaw web interface.
Trying out Gemma 4:12b with OpenClaw
Quite capable, but struggles at times
Now for the real-world tests. I fired up OpenClaw through its web interface and tried a few prompts:
- Create a spinning 3D cube in Python.
- Elaborate on the modern DOOM trilogy, from its 2016 release to The Dark Ages: Revelations.
- Generate a 3D Hello World image.
- Write a short story about a banana trying to enter a Fruits Club, with the banana as the outcast, followed by a moral of the story.
- What would you do if you were a cockroach for a day?
Each prompt had varying responses. The first prompt didn’t produce any meaningful results. It did generate a Python script, but it never really worked. The script just failed to work and bugged out, making it unusable.
As for image generation, that didn’t work out either. OpenClaw is text-based and cannot generate images or videos on its own. However, it did generate a prompt that should (in theory) be used to generate the image, but by using it on yet another agent.
It’s safe to say OpenClaw isn’t designed for heavy-hitting tasks and instead relegates itself to more basic, productivity-focused work. It does, however, work very well with text-based answers.
The DOOM trilogy notes were surprisingly detailed, covering the differences between the games and a bit more about their release — which is something I never told it to do.
OpenClaw can use a variety of LLM models, but it’s practically never going to match the speed or scope of something cloud-based or hosted.
The admittedly meme-worthy banana and cockroach stories worked out well too, but there was one tiny problem. The text was too short, and anything even remotely complicated (involving multiple steps and instructions) would break the chain of thought.
I ran into multiple failed attempts wherein the AI would just get stuck on a particular prompt, essentially “boot looping” itself until I hit the stop button. All of this goes to say that while OpenClaw is a really fascinating local AI solution, it is in no way comparable to proper cloud-based tools.
There’s more than just local models to rely on
Of course, you can always use a cloud-based AI model instead, but it involves a series of steps, including grabbing a unique API key from your host (Anthropic in the case of Claude). Once onboarded and configured, OpenClaw will directly communicate with Claude to get things done.
This makes OpenClaw much more usable and far more capable in what it can and cannot do. That said, cloud-based services will still gut you on token usage, and you also have to consider the privacy aspect. Still, you might want to divert your attention elsewhere too.
