Wren¶

Wren is a small attention-only model that reads bytes and turns a short command into proposed calls on a fixed set of targets, on a CPU. You list the things a command may act on, such as lights, a heating setpoint or a pump, with their current values. Given "close the awning in 10 minutes", Wren answers set(canopy, DOWN) in 10m, with a confidence for each call. The shipped demo answers in about 0.05 s, needs only NumPy and packs into about half a megabyte.

Wren is meant for the commands people give again and again, on one installation. Each installation trains its own model; the demo took roughly 40–60 minutes on a 20-core CPU. Your application executes accepted calls, schedules delays and sends unhandled commands to a larger model. Wren is part of Fluksio automation software, but the library stands on its own.

Quick start¶

You need Python 3.12 or newer and Git. In an activated virtual environment:

python -m pip install git+https://github.com/stroblme/wren
wren "turn on the kitchen light"
# >> set(kitchen_light, true)  1.00

The distribution is named wren-lm and imports as wren. The wren package on PyPI is an unrelated project.

The package ships a demo model trained on a home with 20 targets (wren/home.json), which the command line uses by default. For your own installation, follow the How-to to create home.json with targets, examples and tests, then install the training extra:

python -m pip install "wren-lm[train] @ git+https://github.com/stroblme/wren"

Train and load a pack with Python, or use the tracked CLI workflow:

import json
from pathlib import Path

from wren import Target, accept, act, load_pack
from wren.train import fit

home = json.loads(Path("home.json").read_text())
targets = [Target(**t) for t in home["targets"]]
fit(
    targets,
    out="home.bin",
    examples=home.get("examples", []),
    calibration=home.get("calibration", []),
    tests=home.get("tests", []),
)

model = load_pack("home.bin")
calls = act("close the awning in 10 minutes", targets, model=model)
applied = accept(calls, model=model)

act returns calls without executing or filtering them. Apply your confidence and fallback policy before acting; the How-to covers partial answers and the demo's limits. See the Reference for the API, training defaults and C runtime, or Development for checkout setup, tests and integration. Model architecture, measurements and study reproduction are covered in Research and benchmarks.