v1.6.6 Released: Smarter Local AI, RVC Voice, Skills Builder & More

Hi everyone,

v1.6.6 has been released:

Local AI Model Update:
The base included local model has been changed from Llama 3.2 3B to Gemma 3 4B. The new model is more intelligent and provides a better baseline for local CPU-only use across all 14 available languages in CielChan. Additional improvements have also been made to response generation, allowing Gemma 3 4B to match the response times of the previous smaller Llama model, and in some cases even perform faster.

Quick Chat has also been reworked. It no longer requires a separate model and will now use your normally selected model. Quick Chat is enabled by default for CPU-only users, but is automatically disabled when GPU acceleration or Cloud LLMs are enabled. It can still be optionally re-enabled in those cases if preferred.

RVC Voice Support:
Added RVC voice conversion support. RVC allows speech to be transformed into different voices in real time with very little overhead. RVC is available as a voice option for both CPU-only and GPU users. Select the RVC option from the settings menu to warm up the model and freely use the included HQ Ciel voice model.

This RVC release includes one bundled RVC voice model that can be used across all supported languages.

Additional Progress Indicator During Environment Preparation:
When starting up CielChan, there is now an additional % indicator to help show how much of the startup process has completed.

Skills Builder (Experimental):
Added the new Skills Builder, allowing you to more easily create and manage basic custom assistant tool skills.
Go to Settings > Skills and you will see a Generate button. Select Generate to open the Skills Builder.

Observation & Watch Together Improvements:
Improved Observation and Watch Together modes with better media-aware responses and smoother observation handling. Responses should now feel more contextually relevant to current and recent moments.

Tool Improvements:
Improved assistant tool handling, permissions, and status feedback. Tools work best when used with larger models.

Dynamic skills and tools now integrate more cleanly with the assistant’s tool system. Custom-created skills made through the Skills Builder can also be uploaded to the Steam Workshop.

Onboarding Improvements:
Improved first-run onboarding with a better setup flow, improved language handling, and smarter default model selection across the 3 avatar types.

UI Refinements:
Additional improvements have been made across chat, settings, onboarding, the Skills UI, and assistant controls. Settings has now been redesigned from a single-page menu into themed panels/tabs: Chat & Voice, Seeing & Listening, Companion, Models & Performance, and Advanced. A "Light Mode" settings UI has also been added and can be found within the Advanced tab at the bottom of the page alongside "Reset All" and the other UI-based settings.

Hover panels with additional information now appear when hovering over the vertical sidebar options. These panels can optionally be disabled in the Settings menu.

The UI will continue to receive further improvements over the next few updates.

Performance Improvements:
Further improvements to general response speed and generation efficiency across various parts of the app. Overall CielChan disk space requirements have also been slightly reduced.

CielSocial Improvements:
Added some small improvements to the CielSocial experience, including larger post sizes. Clicking a post will now open a panel allowing you to view the full post without it being cut off in the feed when larger posts are made by users.

A few other small improvements and fixes.

More to come soon.