This article is intended to be a little experiment. I am trying to dictate the text using a tool with Speech-to-Text capability. The tool is called Speed of Sound and is available on FlatHub. The link can be found in the link list below.
Other topics covered are the Project Bluefin Blog, the Hummingbird project, and Red Hat Enterprise Linux’s Lightspeed project. You will also find links for these in the link list. There is one more link: the Linux App Summit. This is a conference that deals with how KDE and GNOME development occurs, or topics like Flatpaks, distributions, or distroless containers.
Link List: * https://docs.projectbluefin.io/blog/bluefin-spring-2026/ * https://fedoramagazine.org/find-out-how-your-fedora-system-really-feels-with-the-linux-mcp-server/ * https://github.com/rhel-lightspeed/linux-mcp-server * https://fedoramagazine.org/fedora-hummingbird-linux-taking-the-hummingbird-model-to-the-full-os/ * https://flathub.org/en/apps/io.speedofsound.SpeedOfSound * https://conf.linuxappsummit.org/event/9/timetable/#20260516
My setup for Local AI consists of LM Studio and Goose. For the Large Language Model, I am using Qwen 3.5 in the 9B variant. I have connected these two tools together. Furthermore, I installed the Linux MCP Server from the Lightspeed project using Homebrew.
The goal of the MCP Server is to assist with system administration tasks, including collecting hardware information, providing an overview of services running, CPU load metrics, and other data points. I have attached a few screenshots.
This marks the end of my Speech-to-Text attempt. For next time, I need to prepare better; with a bit of practice, it could be helpful.
What else is helpful for me: local LLMs. Yes, Speed of Sound also runs with local AI :)
Additionally: generating simple bash or Python scripts, analyzing bank statements. Things that I would never give to Mistral, and definitely not to GPT, etc.
Apart from that, I have no problem talking about finance; quite the opposite—I actually find it important, as it combats envy and usually helps people in other areas too.
But back to the Linux MCP Server, LM Studio, and Goose: Considering this project has only existed for a few months, it is doing a very good job. With the next better local LLM, there will likely be even more room for improvement. I feel that Qwen handles this MCP better than, say, Gemma 4.
One quick comment on the Linux App Summit 2026: I find Flatpak development exciting; otherwise, I am already quite a fanboy of BlueOS/Bazzite and @jorge@hachyderm.io.
The talk at LAS2026 was entertaining and makes me eager for more :)
Personally, Bazzite provides exactly what I want: a stable system that consists 99.9999% of media consumption (gaming, Twitch, YouTube, blogs, music, etc.), and the rest is “work,” like writing a letter or dealing with some financial stuff.
I don't want to worry about dependencies during updates, and it doesn't matter to me if Flatpaks need a bit more RAM or storage space on the HDD. Compared to games that take up 20, 70, or 200GB on disk, this is simply irrelevant. Okay, current RAM prices make one question their necessity, but hopefully, the problem will solve itself over time.
And what about AI on the home server?
ramalama serve --runtime-args="--cache-ram 0 --ctx-size 1024" hf.co/unsloth/Qwen3.5-4B-GGUF:Q4_K_S --port 8120
Ramalama has a default ctx-size of 4096, which causes memory issues with 16GB of RAM and a 4B model—it uses more than is available. With --ctx-size 1024, the RAM usage corresponds roughly to the size of the model download.
I wanted to test the speed; it was expectedly slow, but for system analysis help, it should be sufficient. I initially just wanted to lay a foundation for RHEL Lightspeed; everything else will follow step by step. I still need to read up more and test further on the desktop first.
I am also not yet fully comfortable with ramalama and podman—I need to do more with those too. My OpenCloud and Searxng have been running in podman for over a year now; everything is well automated, and I only had to fix one certificate in OpenCloud. New projects are needed for practice and testing :)
#wiulinuslog #rhellightspeed #rhelhummingbird #localai #bluefin #bazzite #gooseai #lmstudio #las2026