About

I'm Younes. I build software, and I like knowing exactly what my machines can do. This blog is where I write down what local models and agents actually do on my hardware.

How I measure

Every benchmark lists the machine, the runtime and its version, the model build, the settings and the date. I run the same prompts on every build, several times, and report the spread, not just the best run. The raw data is linked from each post so you can check my numbers or run them yourself.

What runs the experiments

A Mac Studio with the M5 Ultra and 96 GB of memory runs the models, the agent and most of my services. The main model is Qwen3.8 Flash Next on sushi, an MLX-based server; oMLX, llama.cpp and Ollama run the rest. Agent runs use Hermes Agent. The full setup.

Analytics

I count visits with Umami, which I host myself. It sets no cookies and builds no profile of you: it records the page, the referrer, your browser and country, and clicks on a few buttons such as share and follow.

Newsletter

If you subscribe, your email address is stored by Resend, which sends the emails for me. You confirm by email before you get anything, every issue has a one-click unsubscribe, and I use the address for nothing else. To have it deleted, unsubscribe or tell me.

AI pulse

The AI pulse is picked and written by Hermes, my AI agent, and published without me reviewing each item. It links to the original source every time. If an item is wrong, tell me and I take it down.

Corrections

If a number looks wrong, tell me. I fix the post, say what changed at the top, and keep the old data.