Repository navigation
AI Playground Version v3.3.0 beta
AI Playground 3.3.0 beta is an update from 3.2.1 beta, with Agent Maker, background agent that runs custom steps, and new image-to-3D and model-fit guidance.
Version 3.3.0 beta updates
- Agent Maker: The Agent preset is available as Agent Maker. Each session keeps its own backend, model, capabilities, and MCP servers. You can save an agent, schedule it, and save it as a Chat preset that opens a session set up the same way.
- Sessions keep running: Switching sessions no longer stops a turn. Chat, Agent Maker, and Home Agent wait in a turn queue. A turn that needs you pauses and says what it is waiting for, including a missing model download.
- Answer agents from Home Agent: An agent can notify the channel you choose (Telegram, Slack, or LAN chat), and can attach an image, a video, or a voice clip. You answer on that channel — reply to the message, reply in the thread, or tap Answer — and the reply goes back to the agent's session.
- Agent results stay in the transcript: Generated images, edits, video, 3D models, and spoken replies stay in the Agent Mode transcript.
- Document search: Agent sessions can search documents indexed for RAG, including several files in one request.
- Pixel3D-Trellis: New image-to-3D preset. It builds a textured 3D model from a single image.
- Models that fit this PC: llama.cpp models show whether the model and its context fit available memory, in the prompt bar and the model picker. A filter keeps models that fit. llama.cpp inference can also be routed to the CPU.
- Shared Use Mode: The setup wizard can turn on shared demonstration use, with image safety set to high. Leaving it needs the Demo Passcode. In that mode, hovering the prompt bar opens the preset picker again; the picker stays hidden only while a tour is on screen.
- Speech: Kokoro reads scaled currency amounts in full. OpenVINO text-to-speech on Windows uses the bundled espeak-ng voice data.
- GPT-OSS MXFP4 downloads: Downloads follow the renamed GGUF files (
gpt-oss-20b-MXFP4.ggufand the singlegpt-oss-120b-MXFP4.gguf). - Security updates: Updated app and Python dependencies, and tightened cloud calls, custom-node installs, and how a shared install folder is chosen. LAN chat passwords can include non-ASCII characters.
Known issues
- Windows SAC (Smart App Control) can cause installation issues with PyTorch and ComfyUI. It is recommended to turn off when installing AI Playground
- Pre-release software: agent, game, cloud, and local Home Agent features may still change — review network and security settings before enabling LAN access
- The installer can fail due to environmental issues. See the Troubleshooting guide for help
- Systems with discrete GPU and integrated GPU may need to disable iGPU when installing
- Systems with more than one GPU may need to disable idle GPUs when inferencing
- An interrupted download can block models from downloading properly. To fix this, go to the model location, delete the temporary file, and try again
- The Colorizer workflow is limited to Xe2-supported GPUs (LNL and BMG)
- A PC restart may be needed after installation for optimal functionality and performance
- Reinstalling over an existing AI Playground installation will delete any ComfyUI custom nodes
- Users can adjust settings beyond their specific system's ability, resulting in failed generations
- Upgrading from 3.1.x may use a different shared model folder layout — verify model paths in Models settings after install
Version 3.3.0 beta supported hardware
- Intel Core Series 3 Processors with 12GB+ of system memory (WCL)
- Intel Core Ultra Series 3 Processors (PTL)
- Intel Core Ultra Series 2 (H) Processors (ARL-H)
- Intel Core Ultra Series 2 (V) Processors (LNL)
- Intel Core Ultra Series 1 (H) Processors (MTL-H)
- Intel Arc B Series GPU Cards (BMG)
- Intel Arc A Series GPU Cards with 8GB+ Memory (ACM)
- Nvidia GeForce RTX GPUs
- Home Agent feature requires a minimum of 32GB RAM on laptops, or 16GB with a discrete GPU