Using SunSys AI Studio

First use

The first time you use Guide Studio, we recommend opening Settings → Guide Studio → Engines and access to install the inference runtimes that match your hardware (CPU, NVIDIA CUDA GPU or Apple Metal). SunSys suggests the recommended variants for your configuration.

For dictation and read-aloud, go to Settings → Guide Studio → Voice: by default SunSys AI Studio uses the operating system’s voices. The section has two tabs, Dictation (STT) and Speech synthesis (TTS) — install local engines there (Whisper, Pocket TTS, Kokoro, Supertonic) for better quality and fully offline processing. Kokoro, Pocket and Supertonic share a single TTS port; only one local TTS engine can be active at a time. Installation can continue in the background — a floating banner and a badge on Settings show its progress.

To enable web search in your conversations, configure SearXNG in Settings → Guide Studio → Tools. MCP servers can also be added there to extend the model’s capabilities. OpenAI-compatible network access is configured in Gateway.

Vision settings (analysis resolution for attached images) are in Settings → Guide Studio → Customisation. Download a model from Settings → Guide Studio → Models before starting a conversation.

Conversations and projects

Create standard or ephemeral (unsaved) conversations. Group them into projects, rename them, duplicate a branch of a discussion or pick a conversation up later from the Guide Studio sidebar.

Active model

The selector at the top of the screen lets you choose and start a model from Settings → Guide Studio → Models. Until a model is running, sending messages is blocked. A green dot at the bottom of the menu shows that a runtime is active.

Attachments and vision

Attach images or documents to your messages. Vision-capable models analyse the visual content. The analysis resolution can be set in Settings → Guide Studio → Customisation.

Dictation and read-aloud

Dictate your messages through the microphone or have answers read aloud. By default the application uses the system voices. In Settings → Guide Studio → Voice, choose a Whisper model (Tiny to Large — it downloads automatically when selected) and a local TTS engine (Pocket TTS, Kokoro or Supertonic), with model preparation and a voice test. In the desktop application, dictation goes through the system or Whisper.

Tools and search

If web search (SearXNG) is configured in Settings → Guide Studio → Tools, or MCP servers are added in the same section, the model can call them during the conversation.

Context compaction

The context ring below the input area shows how much of the token window is in use. Use Compact context to summarise older history and free up room without erasing the visible thread. The threshold and automatic compaction mode are set in Settings → Guide Studio → Performance.

Agentic loop and limits

When the model calls tools (web search, URL reading, MCP…), an agentic loop can chain several turns. A counter shows the budgets used (turns, web, fetch, image, MCP). Per-message caps and safety options (confirmation before an MCP write, pause at the turn cap) are configured in Settings → Guide Studio → Tools.

Guide Studio / Image Studio conflict

Starting a Guide Studio model while Image Studio is running — or the other way round — opens a dialog: cancel, keep both runtimes or close the other one. Running both significantly increases memory use. Turn on Exclusive model launch between studios in Settings → General to close the other runtime automatically on launch.

Setup

Enable Image Studio in Settings → Image Studio and install the sd-server runtime (stable-diffusion.cpp) — separate from the llama-server runtime used for Guide Studio models.

Download at least one diffusion model from the Models tab of Settings → Image Studio. The Image Studio entry in the main menu only appears once the studio is enabled, sd-server is ready and a model is installed.

Visibility in the menu

If Image Studio is missing from the sidebar, open Settings → Image Studio: check that the studio is enabled, that sd-server is installed (and up to date) and that at least one image model is downloaded.

Sessions and projects

Each session keeps its generation history. Organise your sessions into projects to find your work easily and run several iterations on the same theme.

Generation modes

Create a new image from a prompt, edit an existing image (inpainting / guided retouching) or compose from several reference images, depending on the installed model.

Format and resolution

Choose the aspect ratio (square, landscape, portrait…) and the resolution. Presets range from fast (512 px) to 2K for the most powerful machines. You can describe the scene in your own language: SunSys adapts the prompt for the model.

Image models

Select the diffusion model in the top bar. Install new ones from Settings → Image Studio → Models. Some models are specialised (photorealism, anime, image editing).

Conflict with Guide Studio

Image Studio uses its own sd-server runtime. If a Guide Studio model is already running, launching the studio (or the other way round) asks you to cancel, keep both or close the other runtime — running both increases memory use. Turn on Exclusive model launch between studios in Settings → General to skip the prompt.

Guide Studio section

Gathers everything related to Guide Studio. Available tabs: Engines and access, Models, Downloads, Tools/MCP Guide (SearXNG, MCP, agentic safety, call limits), Voice, Performance, Customisation, Mobile link and Gateway.

Engines and access

The Local engines panel brings together the GGUF engine (llama-server), the MLX engine on Apple Silicon, the server port and the LLM models folder. LLM models use a base port (default 18080, then 18081…); the other SunSys services use the 19080+ range (SearXNG, TTS, image). Install or update the runtimes and check they match the SunSys-supported version (green tick, or “?” if the installed version differs).

Models

The Models tab centralises your GGUF and MLX files: browse the disk, consult the SunSys catalogue or the Hugging Face hub, and launch runtimes. Three sub-tabs — SunSys recommendations (default), Downloaded models and Hugging Face — share Choose folder; only Downloaded models offers Scan folder. A hardware summary (RAM, GPU) and an optional Hugging Face token help you choose on the catalogue tabs.

Downloaded models

Lists the GGUF and MLX files in the models folder. Each entry shows the format, vision/MTP/tools badges, companion files, size and an estimated RAM requirement. Adjust the Launch options (vision, MTP) if needed, then start or stop the runtime from the footer.

Catalogue and Hugging Face

SunSys recommendations offers a tested catalogue (GGUF and MLX). The Hugging Face tab provides a full hub search with filters (author, library, pipeline, sort, size). Each variant shows a compatibility badge. For GGUF repositories, download the model alone or the full set with mmproj and/or MTP; for MLX, the complete repository snapshot.

Downloads

The Downloads tab tracks every transfer: direct URL, Hugging Face hub or SunSys catalogue. Each transfer shows progress, speed and time remaining; pause, resume and cancel are available. Multi-file repositories (shards) are grouped. Active downloads are flagged with a badge on Guide Studio.

Tools

Configure SearXNG for web search in Guide Studio, connect MCP servers and set per-message tool-call limits (agentic turns, web searches, URL reads, MCP, image generation). The Agentic safety section can ask for confirmation before an MCP write and offer to continue when the turn cap is reached.

Voice

Organised into Dictation (STT — Whisper) and Speech synthesis (TTS — Pocket TTS, Kokoro, Supertonic) tabs. Kokoro, Pocket and Supertonic share a single TTS port; only one local TTS engine can be active at a time. A background-installation banner lets you return to Voice while installation runs. System voices are used by default.

Performance and customisation

In Performance: context size, max tokens per answer, automatic compaction, GPU acceleration, CPU cores, batch size and the Allow several active models option. In Customisation: system instructions, generation parameters and image processing for vision.

Mobile link and gateway

Mobile link exposes Guide Studio on your local network through a QR code (models on this PC, end-to-end encryption). Gateway exposes an OpenAI-compatible endpoint on the local network to relay requests to your running models.

Image Studio

Two tabs: Settings (activation, sd-server runtime, folders) and Models (catalogue and download of diffusion models). The main-menu entry appears once the studio is ready.

Translation Studio

A workspace dedicated to translation, available from the main menu or Settings → Translation Studio. This feature is under development.

General

The General section gathers cross-cutting options, including exclusive model launch between studios and detailed engine logs.

Advanced

The Advanced section lets you set the URL of the software-resources repository used for runtime downloads (leave empty for official sources; the SunSys software-resources repository is used by default). A test button checks that the repository can be reached from the desktop application.

Notifications

Turn on system alerts for completed Guide Studio answers, pauses awaiting action, or the end of an Image Studio generation. A button sends a test notification.

What the logs contain

Messages from llama-server, the MLX engine, the Downloads manager and the associated services (SearXNG, sd-server / Image Studio, Whisper, Kokoro, Pocket, Supertonic, MCP). Errors and warnings appear first.

Automatic refresh

The view refreshes every few seconds. Use the Refresh button to force an immediate update if you are waiting for a specific event.

Clearing the logs

The Clear button deletes the displayed history (after confirmation). This does not affect the runtime; new messages will accumulate afterwards.

When to check them

When a model will not start (check Engines and access), an answer comes back empty, a download is stuck or an MCP tool cannot be reached, the logs often give the exact cause (port in use, not enough memory, missing binary…). With detailed engine logs turned on, they also help diagnose slow inference.