Hi everyone,
Following testing on the Beta branch, v1.8.2 has now been released to the main branch.
Thank you to everyone who tested the beta and shared feedback. This update includes significant backend changes, interface upgrades, improved multilingual support, reduced resource usage, and further refinements made throughout beta testing.
Backend and Local AI Improvements
CielChan’s base harness and local model framework have been upgraded, allowing newer local models such as Gemma 4 to run within CielChan.
The previous Llama 3.2 3B base model has been replaced with Gemma 4 E2B, which offers improved intelligence, generation speed, memory efficiency, and substantially better multilingual response quality. Startup times and general memory usage have been reduced as well.
The original Llama-based build remains available through the optional Legacy branch for anyone who prefers the previous model and framework.
UI Updates
The main vertical sidebar, chat box, and several other areas of the interface have been updated and reorganized.
A new Home button has been added to the vertical sidebar. Selecting it opens a quick-settings menu containing shortcuts to areas such as:
Activities
Microphone
Performance
Brain
Other commonly used settings
Selecting one of these options takes you directly to the corresponding settings section. This should make it easier for newer users to understand where different features and settings are located.
The UI has received some extra visual polish and expanded color customization. Previously, changing your UI Color primarily affected the vertical sidebar and chat-box header. As of v1.8.2, your selected color is applied throughout the interface for a more consistent appearance.
A new UI Color shortcut has been added to the bottom of the Settings menu for quicker access.
Observation Improvements
Observations now make better use of the context available through your enabled settings. When Vision is enabled, CielChan can better combine current visual information with environmental awareness, recent window activity, conversation context, and your selected observation topic. When Vision is disabled, observations can still use environmental awareness, such as the currently focused application, recent activity, keyboard activity when enabled, and whether you appear to be idle, to provide more relevant comments.
Observation behaviour has been aligned more closely with Watch Together Computer. Comments should now feel more natural and move forward more easily, rather than becoming overly focused on the last chat message, reacting to CielChan’s own previous comments, or making things up.
When little or no meaningful activity is detected, CielChan can shift towards a friendly comment or light question rather than repeatedly describing an unchanged desktop.
Observations are now better grounded in the context that is actually available and are less likely to treat environmental information as something directly seen through Vision.
Vision observations have been improved to further help avoid commenting on CielChan’s own floating avatar, vertical sidebar, or floating interface panels.
Observation Topics
Observation topics are now followed more consistently across standard observations, Vision observations, and Watch Together experiences. When a topic is enabled, it guides what CielChan notices and discusses while still allowing natural reactions to current activity. Without a selected topic, CielChan continues behaving as a general desktop companion and responds naturally using the context available through your enabled settings.
System Audio and Watch Together Improvements
Standalone System Audio observations now use the same general audio-reaction approach as Watch Together. Current dialogue transcription is given clear priority, allowing CielChan to respond more directly to specific things being said instead of giving vague or unrelated reactions. Dialogue coming from media is kept separate from the user’s own speech, reducing the chance of CielChan mistakenly treating dialogue from a video, stream, game, or other media as something the user said.
Watch Together Computer now receives the same direct-dialogue guidance whenever System Audio transcription is available, while continuing to consider relevant computer, environmental, and visual context.
Observation timing and context handling have been refined to reduce missed comments and improve consistency across Vision, System Audio, environmental awareness, and Watch Together combinations.
Ciel Social Improvements
The first automatic Ciel Social post is now scheduled approximately one minute after Auto Post is enabled instead of waiting for the full recurring interval. Later automatic posts remain spaced approximately seven minutes apart.
AI Draft and Auto Post generation now run separately from the main conversation. Preparing a Ciel Social post will no longer clear or interfere with your chat history. While a Ciel Social post is being generated, other background AI processes, including observations, Watch Together reactions, and System Audio commentary, temporarily wait to prevent overlapping requests. Any previously enabled processes resume normally after post generation succeeds, fails, or is cancelled.
Ciel Social now correctly releases its busy state when a draft completes, fails, or is cancelled. This prevents the assistant from remaining incorrectly occupied and allows enabled background activities to continue normally.
Goals and Tool Improvements
Goals now handle research, file creation, editing, inspection, and verification more reliably. Completion is based more closely on what the user actually requested, reducing false failures caused by unrelated tool suggestions or requirements. Verified progress is preserved while a Goal continues, completed actions are not unnecessarily repeated, and saved files are checked before a Goal is marked complete.
Goal results now provide clearer summaries, file locations, and details about anything genuinely left unfinished.
Goal completion and handoff behaviour have been improved. Goals now move cleanly into completed, paused, waiting, failed, or cancelled states instead of remaining indefinitely in “Still working…” or repeatedly trying to finish with no new progress.
Pressing the chat-box Stop button now cancels the active tool operation along with the current AI response, while keeping Tools enabled for future requests. Disabling Tools cleanly cancels any active tool work and running Goal too.
Cancelled or timed-out local AI sessions are now safely released instead of leaving a disconnected session available for reuse. The next request can begin with a fresh connection, reducing repeated no-output timeouts and stalled Goal or tool sessions.
Web Search and Reminder Improvements
Web searches performed through Tools and Goals now provide more consistent results. Relevant details, titles, summaries, and links are kept and presented more clearly once the search finishes.
Reminder requests now understand the requested time and reminder text more reliably across supported languages, including multilingual requests and requests that combine a reminder with other actions.
Movable and Collapsible Panels
The Multiplayer, Book Reader, Observation, Ciel Social, Mini Games, and Focus Companion panels can now be repositioned by dragging their title text. Double-clicking a panel title collapses it into a compact title-only view. Double-click the title again to expand it. Panels moved away from their default position retain their position and open or closed state between sessions. Moved panels can remain open while switching between other parts of the interface, allowing several tools to be arranged around the desktop at once. Panels that have not been moved continue using the existing automatic switching behaviour.
Close buttons have been added to most panels in the top-right corner.
Using the Recenter UI/Model option restores all movable panels to their default positions, expands them, and closes them to return the interface to its standard layout.
Cloud LLM Updates
When using a Cloud LLM instead of a local model, the local model will now be unloaded to free additional RAM as well as some other improvements to Cloud LLM handling.
Other various fixes and improvements to Cloud LLM handling, features and responses.
Avatar Model Settings
Avatar model settings can now be accessed from the bottom of the Companion tab, directly beneath the Personality Cards section.
New Voice Button
A new Voice button has been added to the main vertical sidebar. When additional voices are available for your current language and setup, this allows you to quickly change your AI’s voice directly from the main interface. You can enable the recommended voice settings for your currently selected language from this section as well.
Microphone Button
You can now enable your microphone directly by pressing the Microphone button on the chat-box header bar.
Additional Improvements and Bug Fixes
Tool results, Goal progress, completion handoffs, cancellation behaviour, and local AI recovery have been refined to reduce false failures, repeated attempts, incorrectly running states, and disconnected sessions being reused.
Chat-box activity indicators and placeholder text now return to their idle state correctly when Goals or tools complete, fail, pause, stop, or are cancelled. “Still working…” and loading states should no longer remain visible after the related work has ended.
Various other performance improvements, compatibility updates, bug fixes, and refinements have been included throughout v1.8.2.
Thanks again to everyone who helped test the beta and provided feedback! If you encounter any issues please post them in the discussion boards so they can be investigated.
Build will be live on other platforms and compact branches soon.