Kemudi Read the README
Python · local model

Steering knobs for a local language model.

Each knob adds a direction to the model's internal state while it writes, with no instruction in the prompt. Each one ships with a measurement of what it does to the text and the dose at which the text breaks.

Python · local model
kemudi · previewscreenshot
Kemudi screenshot
2platforms: Python, local model
4knobs, each with a measured usable range
576steered texts scored
6 of 6deliberate code mutations caught by the tests
Why it exists

A knob should come with its measurement.

Kemudi gives a small local language model four steering knobs for a story it is writing: mood, tension, register and speech.

Every knob ships with a measurement: sixteen story openings, nine doses, the same random seed, a random-direction control and a prompt control.

Run it yourself

Start it on your computer.

From the README’s “Start it” section. Full options, requirements and troubleshooting are there too.

Open the README
shell
$ git clone https://github.com/juw12/kemudi && cd kemudi
# Windows: double-click "Start Kemudi.cmd"    macOS or Linux: ./start-kemudi.sh
# or by hand, with Python 3.10 or newer:
$ pip install torch --index-url https://download.pytorch.org/whl/cpu
$ pip install -e .
$ kemudi serve --open
How it works

How it works

  1. 1

    Find a direction

    Sixteen story openings are each written under two opposite style instructions.

  2. 2

    Check it on held-out sentences

    Each knob has 32 hand-written sentences.

  3. 3

    Steer

    While the model writes, a hook adds the vector to the output of that block.

  4. 4

    Measure

    The effect, the perplexity and the repetition of every text are computed by the same model with all knobs at zero.

What's inside

What the four knobs do.

Four knobs

Mood (bleak to hopeful), tension (calm to tense), register (casual to formal) and speech (narration to dialogue), for a story the model is writing.

They act while the model writes

Move a slider and the next words change. Nothing is added to the prompt and nothing is regenerated.

A green band on every slider

The doses that were measured to keep the text fluent. Past the band the app warns you before you get there.

A gauge per knob

It reads the model's own state along that direction, word by word.

Compare with zero

The same opening and the same random seed, written with and without your knobs, side by side.

Measured, with two controls

Each knob has a dose-response with 95% intervals, a random direction of the same size, and a prompt control measured the same way.

It runs on your computer

The model is Qwen2.5-0.5B-Instruct (about 1 GB, Apache-2.0), on the CPU. No account, no API key, no data leaves the machine.

Your data

Runs on your computer

  • The model runs on your CPU.
  • The first start installs Python packages and downloads the model once, about 1.3 GB together.
Good to know

Honest limits.

Straight from the README, so you know what you're getting before you try it.

  • Everything here was measured on Qwen2.5-0.5B-Instruct.
  • The sample is small: 16 openings, 48 tokens each, one random seed, English, story continuation only.
  • The judge is the model itself.
  • Speech is the weak knob: inside its usable range the effect is small.
  • Each knob was measured alone.
Under the hood

How it's built.

Built with

Language
PythonHTML/CSS
Testing
Automated tests

44 files in the repo · last updated 17 Nov 2026.

Browse the code

Recently shipped

  1. One command to reproduce, a dose table with intervals, a start script for macOS and Linux

  2. Keep line endings stable: LF everywhere, CRLF for the Windows start script

  3. Measured steering knobs for a local language model

Full history on GitHub
Questions

Good questions.

What does it run on?

Python, local model.

How do I get it running?

The commands in the “Start it” box above come straight from the README. Paste them into a terminal; the README covers options and troubleshooting.

Where does my data go?

The model runs on your CPU. The first start installs Python packages and downloads the model once, about 1.3 GB together.

Is this the same as Spaceship?

No. Spaceship is an interface study: its knobs are written into the prompt and it calls hosted models.

Does this show that vectors beat prompts?

No. The prompt control failed on this 0.5B model because many of its texts are the control instruction handed back.

What do I need to run it?

Python 3.10 or newer. The first start downloads Python packages and the model once, about 1.3 GB together.

What doesn't it do yet?

Everything here was measured on Qwen2.5-0.5B-Instruct. The sample is small: 16 openings, 48 tokens each, one random seed, English, story continuation only.

Is it free?

The source is public on GitHub under the MIT licence.

Keep exploring

More from the same workbench.

Turn a knob and the next words change.

Steering knobs for a local language model.