wyat

Decisions, made locally.

wyat.io · wyat-1-800m · Documentation

wyat chooses between options your application defines. You provide a message and a set of candidates. A small local model returns the ID of one of them. The model was trained on French and English decisions. Nothing leaves your machine, and nothing runs until you act on the result.

The model is wyat-1 800M, using the PyTorch runtime on CPU or CUDA. Version 0.1.1 downloads its pinned release and verifies every file before loading.

Try it

No model required. The demo uses fixed scores to show the interface.

git clone https://github.com/teilomillet/wyat && cd wyat
uv run python examples/quickstart/main.py --demo
# Selected: cancel

Install

Python 3.12 or newer, with uv.

uv add wyat --extra inference   # PyTorch, CPU or CUDA

Add --extra server to serve over HTTP. The core alone (uv add wyat) has no dependencies. Model weights are separate from the package. wyat.load() downloads and verifies wyat-1 800M on first use, then reuses the verified cache. Run uv run wyat download to fetch it ahead of time. Inference itself never uses the network.

On Apple Silicon, 800M uses PyTorch CPU. The mlx extra supports other native artifacts; it does not support the 800M artifact format.

Use it

import wyat

model = wyat.load("teilomillet/wyat-1-800m")  # also the default for wyat.load()

decision = model.decide(wyat.Request(
    query="Choose the requested booking action.",
    context="Please cancel my booking.",
    candidates=[
        wyat.Candidate(id="create", text="Create a booking"),
        wyat.Candidate(id="cancel", text="Cancel the booking"),
    ],
))

decision.selected  # one of your IDs, such as "cancel"

Load once. Decide many times. The same request and result work in Python, JSONL, and HTTP:

uv run wyat decide teilomillet/wyat-1-800m < requests.jsonl > decisions.jsonl
uv run wyat serve  teilomillet/wyat-1-800m

Besides choosing an option, wyat rates on an ordered scale (Request.score) and answers yes or no with a probability (Request.noul).

Documentation

Coding agents can start from llms.txt.

License

Code is licensed under Apache 2.0; see NOTICE. Model artifacts carry their own licenses. The 800M trained adapters and decision head are released for non-commercial use; its backbone and runtime use Apache 2.0. Contributions are welcome; see CONTRIBUTING.md.