December 31, 2025, 8:28pm 1

Last month I splurged on this laptop, got the version with 128 GB RAM to be able to run decent sized AI models locally.

My first try was to install CachyOS but the results were mixed. Installer crashed randomly, USB devices would disconnect, random crashes, suspend not working, power usage not very optimal, and a few more issues.

Also, for the AI stack AMD ROCm stack is only officially supported for Ubuntu 24.04, but I bit the bullet and installed Ubuntu 25.10 and got the AI stack running.

Now, for the cool part, I was actually getting somewhere with Agent-Zero and using gpt-oss-20b (as utility model) as well as Qwen3-Coder-30B-A3B-Instruct (as reasoning model). The reasoning model could run with 64k context window and the utility model with 32k context window and both models would just fit into the VRAM.

The Ryzen AI MAX 395 is kinda slow in prompt processing, which holds back the performance a bit. I do hope that maybe optimized drivers will improve this over time a bit. Token generation is pretty good around 15 tokens/sec for Qwen3-Coder and more than 20 tokens/sec for gpt.

I had to work on the system prompts a bit, build some memory, but after a while now A0 is able to work very efficiently. For example, I gave it this prompt:

“Download the most recent linux kernel release candidate, download the newest Ubuntu kernel available, use the config from that kernel, configure the kernel with this config and sane defaults for new options, then find and patch the error messages for amdgpu about “LTTPR count is nonzero but invalid lane count reported” and nvme “using unchecked data buffer” to not appear during boot. compile the kernel and create .deb files.”

It took some tweaking and a few memories to get it unstuck at two places, but in the end after about 2 hours it actually produced the correctly compiled deb kernels which then successfully booted and had the nuissance errors removed from the log.

I can see A0 being a real big help. I have added BrowserMCP, but want to change to playwrightMCP so that it can perform complex agentic web tasks. BrowserMCP is limited as it cannot access the DOM model directly in the browser.

This all was on Ubuntu. Wanted to share what I work with and towards.

Does anybody else have the same laptop and maybe got it running well in CachyOS? If so, would love to compare notes. If not, I will probably wait a bit more and then in a few weeks make an effort to try CachyOS again and work to resolve all the issues.

I plan to write up my configuration and guide to getting A0 to run on this hardware for others once I have all the kinks ironed out.

Depending on what you want to do. The Asus ROG Z13 is a laptop. If you want a desktop, there are Framework Destop or GMKTec for less money and a little bit higher TDP.

32GB VRAM is a bit on the tight side for serious AI use, you will have to consider not only the model size but the context window as well. For example Agent Zero will not really work with less than 32k context size in my experience. So consider 2 of these cards if you want high inference performance without offloading to RAM.

In my company we went the opposite way and purchased a dual EPYC server with 24x64GB = 1.5 TB of DDR5-5600 memory. Luckily we purchased this summer before the DDR5 prices exploded.

We use it to run very heavy models like Deepseek-V3.2 670b with 130k context. We get 5 tokens/s which is usable. GLM-4.6 also 5 tokens/s but with 200k context.

There is a distinct difference between smaller models <100b and the >400b models. We run only Q8 quants.

Deepseek 670b with big context can do some really amazing things when it comes to difficult tasks. For example, we made it plan, draft, implement and security-audit a Rust based Websocket backend. Prompting is king, you have to know how to steer it correctly.

That said, it basically delivered a high-quality, production ready backend within 1(!) prompt. I found no error or significant design weakness. The code compiled on first try.

It produced 22 pages of output, ran for 1 hour and 40 minutes for that single prompt. The output included a documentation, and a github style README how to compile and implement.

A good programmer would have taken several hours to solve that. So yeah, for quality work big models with big context are king.

1 Like

Hello and welcome,

Sorry, but this forum is for CachyOS users.

The Ubuntu forum seems to be over here → https://discourse.ubuntu.com/

The guy’s original post spent 90% of the time talking about how he got an AI stack to work on an unofficial version of Ubuntu.

I’m mentioning how I had difficulty getting it up and running on a “supported” version of linux. I’m asking if he has any tips or steps to get it to work on an unofficial version/flavor of linux that he can share.
I’m on a CachyOS forum. The assumption is that I would be applying it to CachyOS, otherwise I wouldn’t be here.

1 Like