@ondeinference/cli

@ondeinference/cli

A terminal app for your Onde Inference account, plus local model fine-tuning.

Install

npm install -g @ondeinference/cli

npm installs the right native binary for your platform automatically.

It works on:

• macOS (Apple Silicon and Intel)
• Linux (x64 and arm64)
• Windows (x64 and arm64)

Other ways to install

Method	Command
Homebrew	brew install ondeinference/homebrew-tap/onde
pip	pip install onde-cli
uv	uv tool install onde-cli
Dart pub	dart pub global activate onde_cli
.NET tool	dotnet tool install --global Onde.Cli
Cargo	cargo install onde-cli

Run it

onde

That opens the terminal UI.

From there you can:

• sign up or sign in
• create and manage apps
• assign models
• fine-tune supported local models
• export merged models to GGUF

No browser needed.

Basic keys

Key	Action
Tab	Move between fields
Enter	Submit or confirm
Ctrl+L	Go to sign in
Ctrl+N	Go to create account
Ctrl+C	Quit

Fine-tuning

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

Training runs locally:

• Metal on Apple Silicon
• CPU on other platforms

So yes, no cloud training setup and no Python environment to babysit.

If you want a quick mental model for what the network is doing once it starts running, Onde has a short write-up on the forward pass.

Supported base models

Model	Size
Qwen/Qwen3-0.6B	~1.2 GB
Qwen/Qwen2.5-1.5B-Instruct	~3.0 GB
Qwen/Qwen3-1.7B	~3.4 GB
Qwen/Qwen3-4B	~8.0 GB

Only safetensors models can be fine-tuned. GGUF models are already quantized, so their weights are not differentiable.

Training data

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|>"}

Running a fine-tune

• Open the Models tab.
• Pick a safetensors model with ↑ / ↓.
• Press f to open the fine-tune config.
• Set your data path, LoRA rank (default 8), epochs (default 3), and learning rate (default 0.0001).
• Start training.

A rank-8 adapter for the 0.6B model is about 1.5 MB, so the output stays pretty small.

After training

• 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 inference.

What is Onde?

Onde Inference is for running LLMs on the user's device. No server round-trips, no sending prompts off to somebody else's machine.

It ships as native SDKs for:

Swift · Flutter · React Native · Rust

The CLI is for account management and local fine-tuning. The SDKs are what you ship in your app.

Debug

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

License

Dual-licensed under MIT and Apache 2.0.

Copyright

© 2026 Splitfire AB (Onde Inference).