bananamend
bananamend runs the BananaMind-2 chat models on your own computer. The models are small: Nano is 40 MB, Mini is 101 MB, and Pro (Preview) is 556 MB.
The engine is Rust. It reads the published model.safetensors and
tokenizer.json directly, so there is no conversion step and no second copy of
the weights. PyTorch is not necessary at any time.
Greedy decoding gives the same tokens as Hugging Face transformers. That
agreement is a test in this repository, and not only a statement
(tests/test_parity.py).
Three ways to use it
| Product | Language | Installation | Use it for |
|---|---|---|---|
bananamendr |
Rust | cargo install bananamendr |
a command line program, and a library for your own Rust code |
bananamendy |
Python | uv pip install bananamendy |
model downloads, a command line program, and a server with the OpenAI interface |
| the browser build | WebAssembly | nothing to install | the demonstration page |
The division of work is deliberate: bananamendr takes a path and nothing
else. It makes no network request. All of the convenient parts — downloads, a
cache, a configuration file, a server — are in bananamendy.
Start here
uv pip install bananamendy
bananamendy pull nano # about 40 MB into the Hugging Face cache
bananamendy chat --prompt "Name one ocean."
bananamendy serve # OpenAI interface on 127.0.0.1:8377
Then point any OpenAI client at the server:
curl http://127.0.0.1:8377/v1/chat/completions \
-H 'Content-Type: application/json' \
-d '{"model": "nano", "messages": [{"role": "user", "content": "Hi"}]}'
The demonstration page
The demonstration page runs the same engine in your browser. The page downloads the Nano model from Hugging Face, and then all of the work stays on your computer. No text goes to a server.
Speed
Nano writes approximately 500 tokens each second on an Apple M4 Max, with all of the cores. The browser build has one thread, and it is therefore slower. The larger models are slower in proportion to their size.
Measure your own computer:
bananamendy bench
bananamendy bench --name mini