About

Hi, I’m Orkun Soylu. I spent 28 years running enterprise IT, and in June 2026 I left the 9-to-5 to chase what I’d been dreaming about: working on AI full time. I build open-source AI tools, and I contribute back to the open-source projects I use wherever I can. I run the current models my projects depend on in my own environment, and compare them.

Why this exists

Keeping up with how fast AI is moving takes a huge amount of effort. A solution goes “old” within months. On the local side of AI, continuous improvement makes a big difference: you can raise intelligence while cutting cost, and raise throughput while improving security. More than half my time goes into testing new technologies, models and what they can do, and, once I find something better, rolling it out across all my past projects. The rest goes into new projects. I run many of my processes with my agents. Every result along the way is valuable, the successes and the failures alike. This site exists to archive the results of my tests. My agent runs the site, but every page is published only after I approve it. I believe that learning from the past and drawing conclusions from tests takes a good archive. You’re on that archive right now.

What’s here

Two kinds of posts, tied together by tags.

Evals. As I said above, I test open-weight models on my own GPUs, on real work: speed, quality, tool use, power. Every eval page lists the date, setup and settings. The failures stay in, too. :)

Writing. On AI, technology, running IT, and what I learned by doing.

New posts arrive by RSS, and on X.

Hardware

Radeon Pro W7800 48 GB, RX 7900 XTX 24 GB, GeForce RTX 4070 Ti SUPER 16 GB. On bare metal and in virtual machines. Every eval is published with the environment it ran in. How I measure is explained here.

Software

I’ve been a Linux user for 25 years. Mostly Debian; on servers I sometimes install Ubuntu. I prefer GNOME, with my own extensions, themes and icon sets. Linux has been my primary operating system since 2019.

Open source

What I build is open source and on GitHub. I support the open-source projects I rely on.

Independence

No sponsors. No vendor has given me hardware, models or money.

Reading the numbers

A number holds for the setup, versions and date on its post. Runtimes and models move fast. When two posts disagree, the controlled comparison wins.