All seven roles

AI models

The brains

The part that thinks. I have run more than ten of them, from small and quick to very large.

An AI model is a program that has read an enormous amount of text and can now read, write and reason. Its size is counted in “parameters”, written with a B for billions. Think of it as the number of adjustable connections in the brain. More usually means smarter, slower, and much hungrier for memory.

Most people who try this at home run models of 4 to 16 billion. My laptop has 128 GB of memory, which let me run models of up to 120 billion on my own desk, with nothing sent to anyone.

8 to 120

billion parameters: the range of models I have run on one laptop

The ones I have worked with

  • gpt-oss 120B

    120 billion · local

    The largest I run on my own machine. It takes about 63 GB of memory by itself. I use it for hard reasoning and for coding by hand.

  • Qwen3 Coder Next

    80 billion · local · retired

    A coding specialist. I retired it after real work showed it was the slowest of three candidates and its size bought nothing extra.

  • DeepSeek V4 Flash

    about 91 GB on disk · local · retired

    A very large open model. It ran, but stacked on everything else it exhausted the machine’s memory during testing. Too big and too slow for daily use, so it went.

  • Gemma 4, my own agent build

    31 billion · local

    For weeks the everyday brain of my assistant. It scored 100 out of 100 in a test shaped like real use, where the model before it scored 70. Today it is the one that can look at pictures.

  • Qwen3 and Qwen3 Coder

    30 billion · local · retired

    Earlier everyday brains and candidates. One answered with invented tool commands instead of really looking things up, which is how it lost its place.

  • “Libero”

    27 billion · local

    A model with fewer built-in refusals, kept only on my own machine. It is switched on by request, for one conversation or a set time, and then the normal model comes back automatically.

  • gpt-oss 20B

    20 billion · local

    The everyday local brain since late August. Chosen after a head-to-head test on real tasks.

  • Qwen3 8B

    8 billion · local

    The quick one. It answers simple messages in about 0.7 seconds, where the bigger model needs 2.

  • An embedding model

    local

    Not a talker. It turns text into numbers so that the archive can find notes by meaning, not just by matching words.

  • Claude, GPT and Gemini

    from outside companies

    Much larger models that I reach over the internet for the hardest jobs and for a second opinion. My main chat assistant currently answers with one of these, with a local model as its backup.

What running them taught me

  1. 1

    The biggest model is not the best one if it does not fit. A brain that pushes everything else out of memory helps nobody.

  2. 2

    Choose by testing on real work. More than once the model that looked best on paper lost.

  3. 3

    One brain cannot do every job. I tested using a single model for both quick and hard messages, and it was worse at both.

Next

The workers

A brain that has been given a goal and permission to act.