Hunting for the frontier models on local boxes
Setting up a 125B MoE model to run at 45–60 tokens per second on a 2019 desktop with a 16 GB graphics card. Experiments with Strata and Qwen3.8-flash-next. Running Claude Opus 4.7 at home.
I’m Jussi, a Finnish software engineer based in Arnhem, the Netherlands. I’m Head of Engineering at Roadie, and through Ilities I help teams untangle codebases and set up the structure that makes AI tools useful instead of noisy. In my own time I build Viberglass, an open-source workspace for teams and their coding agents.
Setting up a 125B MoE model to run at 45–60 tokens per second on a 2019 desktop with a 16 GB graphics card. Experiments with Strata and Qwen3.8-flash-next. Running Claude Opus 4.7 at home.
In software world we can try to slim down career progression of developers into two different tracks, a technical track and a business driven track. Which track one chooses usually comes down to…
During my life I've lived in five different language areas. On two of these I have been able to communicate with the natives very easily because we have been able to speak the same language. On…
Sometimes we have the need to create a proper shape for our endpoints. Since the last year or so I have been working more on the frontend side of a JVM/Spring application I've grown frustrated…
Codebase rescue, architecture reviews and AI-ready engineering practices, through Ilities.
How I can help