Installation

Python: the complete set

uv pip install bananamendy

pip install bananamendy also works. The command installs bananamendr, which is the engine. The wheel needs no compiler, because one wheel serves all CPython versions from 3.9.

Make sure that it works:

bananamendy registry
python -c "import bananamendr; print(hasattr(bananamendr, 'Model'))"

Rust: the command line program only

cargo install bananamendr
bananamendr --version

This program takes a path to a checkpoint. It downloads nothing. Get a checkpoint with bananamendy pull, or with scripts/fetch_models.sh in the repository.

For your own Rust code:

[dependencies]
bananamendr = "1"

The default features read files and use all of the cores. A build for a browser switches both off:

[dependencies]
bananamendr = { version = "1", default-features = false, features = ["wasm"] }

From this repository

git clone https://github.com/twardoch/bananamend
cd bananamend
./install.sh

The script installs the bananamendr program into ~/.local/bin, and it installs both wheels into the interpreter that uv finds as the system Python. It tests the wheels in a temporary environment first, and it makes sure that it changes no file in the repository.

Other locations:

./install.sh --prefix /usr/local --python /opt/homebrew/bin/python3.12

For development

uv venv --python 3.12 .venv
source .venv/bin/activate
uv pip install -e ./crates/bananamendr-py -e ./bananamendy pytest httpx
python -m pytest bananamendy/tests -q

Get the models

Alias	Hugging Face repository	Size
nano	BananaMind/BananaMind-2-Nano-Chat	40 MB
mini	BananaMind/BananaMind-2-Mini-Chat	101 MB
pro	BananaMind/BananaMind-2-Pro-Preview-Chat	556 MB

bananamendy pull nano
bananamendy models          # what is in the cache now
bananamendy where nano      # the directory of one model

The weights go into the usual Hugging Face cache. HF_HOME and HF_HUB_CACHE control the location, and a model that you already have is not downloaded again. Any repository name works, and so does a path to a directory:

bananamendy info --name /path/to/my-checkpoint
bananamendy chat --name SomeOrg/Some-Chat-Model --prompt "Hello"

The engine reads config.json, model.safetensors, tokenizer.json and, if it is present, tokenizer_config.json. All of the tensors must be F32. The engine refuses a different type; it does not convert it.

Configuration

bananamendy init_config     # write the TOML file
bananamendy config          # show the values in use, and the location of the file

platformdirs decides the location. On macOS it is ~/Library/Application Support/bananamendy/config.toml. On Linux it is ~/.config/bananamendy/config.toml.

Each value also has an environment variable. The name is the field in capital letters with the prefix BANANAMENDY_:

BANANAMENDY_MODEL=mini BANANAMENDY_PORT=9000 bananamendy serve

The order is: the value that you give in the command, then the environment, then the TOML file, then the value in the program.