onde

A terminal app for managing Onde Inference apps and fine-tuning models locally.
Sign in, assign models, and train adapters without bouncing between browser tabs.

Install

pip install onde-cli

That installs the native onde binary for your platform. No compiler setup, no Node.js, no extra runtime to babysit. For package docs and the full install matrix, see https://ondeinference.com/cli.

What you get

Run:

onde

and you'll get a full terminal UI where you can:

• sign up or sign in
• create and rename apps
• assign models to apps
• fine-tune supported Qwen models locally
• merge adapters and export GGUF files

If you prefer doing this work in a terminal instead of a browser, that's the whole point.

Keyboard shortcuts

Auth screen

Apps list

App detail

Model picker

Fine-tuning

Fine-tuning

onde can fine-tune Qwen2, Qwen2.5, and Qwen3 safetensors models with LoRA.

Training runs locally:

• on Metal for Apple Silicon
• on CPU everywhere else

No cloud training job. No Python environment. No notebook setup spiral.

If you want the short version of what happens inside the model at inference time, Onde also has a plain-English note on the forward pass.

Supported base models

Only safetensors models can be fine-tuned. GGUF models are already quantized, so this pipeline doesn't use them as training inputs.

Training data format

Use one JSON object per line. Each object needs a text field containing the full conversation in Qwen's chat template.

{"text": "<|im_start|>user\nWhat is the boiling point of water?<|im_end|>\n<|im_start|>assistant\n100°C at sea level.<|im_end|>"}
{"text": "<|im_start|>user\nWhat about at high altitude?<|im_end|>\n<|im_start|>assistant\nLower, around 90°C at 3000 m.<|im_end|>"}

Running a fine-tune

1. Open the Models tab.
2. Pick a safetensors model with ↑ / ↓.
3. Press f.
4. Set your training data path, LoRA rank, epochs, and learning rate.
5. Start training.

Default values:

• LoRA rank: 8
• Epochs: 3
• Learning rate: 0.0001

As a rough reference, a rank-8 adapter for the 0.6B model is about 1.5 MB.

After training

Once the run finishes:

• press m to merge the adapter into the base weights
• press g to export the merged model to GGUF

The exported file loads directly in the Onde SDK for on-device AI inference.

Other ways to install

Build from source

git clone https://github.com/ondeinference/onde-cli
cd onde-cli
cargo build --release
./target/release/onde

Platform support

Prebuilt native binaries are available for:

Debug logs

Logs are written to ~/.cache/onde/debug.log.

The app uses the terminal screen directly while the TUI is open, so logs go to the file instead of printing over the interface.

Related SDKs

Source and issues

This package ships a prebuilt native binary. The source code lives here:

github.com/ondeinference/onde-cli

Bug reports and feature requests should go there too.

License

Dual-licensed under MIT and Apache 2.0.

• MIT License
• Apache License 2.0

Copyright

© 2026 Splitfire AB (Onde Inference).