Laya local setup

The upstream Laya project publishes Python code and model weights for local inference. This guide points to the maintained install path and explains where this site's hosted API fits.

Choose where to run it

Install the upstream Python package

The project currently lists Python 3.10 or newer. Use a virtual environment so the install and your program use the same interpreter. The repository has separate Windows instructions and current dependency notes.

python3 -m venv .venv
.venv/bin/python -m pip install laya
.venv/bin/python -I -c "import laya; print(laya.__version__)"

The version check confirms the package import without loading a model. Loading a checkpoint requires access to its published weights on the first download. Review the current upstream installation guide ↗ before using these commands in production.

Find the model weights

The upstream English and multilingual checkpoints are published on Hugging Face. Laya AI's hosted service offers these two choices as laya-english and laya-multilingual. The upstream package may expose additional checkpoints and routing modes; check its documentation for the version you install.

• English checkpoint and model card ↗
• Multilingual checkpoint and model card ↗

Apple silicon and MLX

An independent Laya-MLX port ↗ publishes a native Apple silicon runtime and converted weights. It is maintained separately from the upstream Python package and from this hosted service. Check its requirements and model cards if local Mac inference is your goal.

Test a decision before operating a model

Use the playground to try a short state and a typed question, inspect the full answer, then decide whether a hosted API or your own runtime matches your deployment needs. Results on your own labeled examples matter more than a generic latency claim.