[Bug/Feature] Flutter SDK fails with Supabase SocketException and lacks documentation for custom/local model loading

[RaihanulHaque]

Description
Hi there! I’ve been using the Cactus CLI to run a finetuned Qwen 3 0.6B (finetuned with Unsloth and converted via cactus convert). It works flawlessly in the terminal.

However, I’ve hit a wall when trying to move this into a mobile app using the cactus_flutter SDK. I'm facing two main issues: a hardcoded network error and a lack of documentation for custom models.

---

1. Hardcoded Supabase SocketException

When I run the downloadModel() function from the basic_completion.dart example, the app crashes with a SocketException. It seems the SDK is trying to reach a specific Supabase instance that is either private or unavailable.

The Error:

I/flutter (12407): Error fetching models: SocketException: Failed host lookup: 'vlqqczxwyaodtcdmdmlw.supabase.co' (OS Error: No address associated with hostname, errno = 7)

Even if I have a stable internet connection, the app cannot resolve this host, which prevents the model initialization flow from completing.

2. No path to load custom/finetuned models

The current Flutter implementation seems built around downloading a pre-set list of models from your backend. Since I have a custom finetuned model (ryuma007/qwen3-0.6B-hishab_json), I cannot find any documentation or API methods to:

• Point the Flutter SDK to my own Hugging Face repo.
• Load the converted model files directly from the local device storage (after transferring them manually or via a custom download).

---

Request for Help/Fix

1. Fix the Supabase Dependency: Can we have a way to bypass the model-fetching logic if we already have the model files or a direct link?
2. Custom Model Documentation: Could you provide an example or a method (e.g., lm.loadFromPath) that allows us to use models converted via the CLI rather than just the ones in the default gallery?

My Setup:

• Model: Qwen 3 0.6B
• Tools: Unsloth for finetuning -> Cactus CLI for conversion.
• Platform: Flutter (Android)

Thanks for the great work on the inference engine, hope we can get the Flutter side working just as smoothly as the CLI!