OB Studio

OB Studio · a project in progress

I'm training a small language model from scratch, to run on my own laptop.

It is called Tot-360M. It is the model behind Tot, an AI companion I am building. This site is where I keep an honest record of the work, including the runs that failed.

Read about Tot-360MTry the Tot desktop app (alpha)Try the Drift browser (alpha)

Why build a model instead of renting one?

Tot started on other companies' models, reached through OpenRouter. That works, but the assistant is only as private, cheap and steady as someone else's service. The direction now is a model I train myself, small enough to run on a laptop and a Raspberry Pi, so everyday conversation and simple actions can stay on my own hardware.

Training still happens on rented cloud accelerators. “Local” describes where the finished model is meant to run, not where it is made.

Tot is the product and Iris is the personality I am aiming for: soft, understanding, quietly determined, and willing to disagree honestly. When someone is overwhelmed it should give one manageable step, not a list of ten.

What exists today

These are things I have built and checked. None of them is a benchmark score, because there is no trained model to score yet.

Parameters
345,359,296
Counted from the actual architecture, with shared word embeddings. Named 360M for short.
Tokenizer
32,000 tokens
Built for this model. It turns text into the numbers the model reads.
Training data so far
~7.2M tokens
A calibration snapshot in 16 verified files with held-out examples. It is not the full pretraining set.
Architecture
Qwen3-compatible layout
Trained from scratch. Not a renamed pretrained Qwen. The layout matches what llama.cpp already runs.

Where it stands right now

The foundation is tested. The first real training run is still waiting for a TPU.

Progress snapshot · October 10, 2026, 18:04 KST

  1. 01

    Architecture and PyTorch implementation

    Parameter count and tied embeddings checked without allocating the full weights. Attention, gradients and loss behaviour pass numerical checks on a two-layer version.

    Done
  2. 02

    Tokenizer and calibration data

    32,000-token tokenizer and about 7.2 million packed tokens, with separate held-out examples.

    Done
  3. 03

    Export and inference compatibility

    A miniature model exports and runs on a pinned llama.cpp build. A miniature model also saves its training state and resumes in a fresh process from real data files.

    Done
  4. 04

    Kaggle TPU training run

    Version 4 got a TPU and verified the inputs, then failed starting the eight parallel workers because Kaggle set a conflicting single-worker address. No training step was verified. Version 5 has the fix and is submitted. At the snapshot time it was still queued for a TPU.

    In progress
  5. 05

    Full pretraining

    Not started. Needs a working distributed run first, then the full dataset.

    Not started
  6. 06

    Personality training and evaluation

    Iris personality training, quality evaluation, and measured speed and memory on a laptop and a Raspberry Pi. The earlier RAM and speed figures were targets, not measurements.

    Not started

The first goal is capability in the range of Qwen3-0.6B, within roughly five points on the Artificial Analysis Intelligence Index, with useful conversation and simple tool use. That is an ambition for version one. Tot-360M has no measured score yet.

Other things I've made

All projects

Browser games, simulations and small tools from before the language model. They still run.