Open media
Lailama brings the power of artificial intelligence to HomeGenie using language models that run entirely on your own computer. Directly from the AI chat, you can ask questions about your home, control your devices, adjust settings, and build custom automations or dashboard widgets.
Because the AI runs locally on your machine, your data never leaves your home network. Once your chosen model files are downloaded, the assistant can understand and control your home completely offline. Any integrations that depend on external online services will still use their standard internet connections.
Your home, your private local assistant
Like Gemini Genie, Lailama is an AI Genie connected to HomeGenie's core AI services. Both share the same underlying architecture to inspect devices, discover configurable options, and execute permitted actions. The fundamental difference is privacy and autonomy: Lailama runs local GGUF models directly on your hardware via LLamaSharp.
You can start talking to your assistant with everyday requests:
- “Are any exterior doors open, and which lights are currently turned on?”
- “Turn off all downstairs lights and set the porch light to 40%.”
- “What is our current household power consumption, and are any high-load appliances running?”
- “Which program controls the motion sensor for the porch light, and what is its current shut-off delay?”
- “Change that shut-off delay to two minutes.”
When the necessary capabilities are enabled, you can also ask Lailama to generate automations and dashboard widgets. Keep in mind that performance depends on the model you select: a lightweight model that excels at device switching may need more guidance when writing complex multi-step code.
Setting up Lailama
Getting started is simple. Once the Lailama package is installed and enabled from System settings, you can set it up directly in the chat:
- Open the AI chat and select Local AI from the assistant list.
- Click the gear icon to open settings (or access them via System settings).
- Choose an AI Model from the curated catalog. Status badges help you pick the right one, including Recommended, Tested, Limited, or Experimental.
- Save your settings. HomeGenie will automatically download the necessary model files. Please wait a moment while the model downloads, loads into memory, and completes its warm-up phase.
- Once the assistant is ready, try a simple opening question like: “Which lights are turned on?”
The initial setup may take several minutes depending on your internet connection and the size of the model. Loading weights into system memory and preparing the context cache are separate from the initial download; subsequent restarts will be much faster because the model file is already saved locally.
If a model fails to load, the chat interface displays an informative error message instead of staying stuck in a loading loop. Check your notifications and backend logs for details. Some GGUF files use experimental tensor structures not yet supported by the local LLamaSharp runtime; switching to a tested model from the list will resolve the issue.
Choosing the right model for your hardware
The built-in model catalog is tested and ranked specifically for HomeGenie. When you open the AI Model List editor, you will find practical requirements and test notes for each model file. These notes reflect real-world observations from our test environments:
- Qwen3-4B-Instruct-2507 is marked as Recommended: it delivers reliable device control, accurate parameter reading, and balanced response times.
- Qwen3.5-9B-DeepSeek-V4-Flash is Tested: it handles complex parameter discovery and multi-step actions with high precision, but requires more RAM and processing time.
- Qwen3-1.7B is marked as Limited: it handles direct switch commands and brief summaries quickly, but can be less reliable with complex feature discovery or nested instructions.
Use these recommendations as a helpful guide for your own setup. When trying a new model, start with simple status questions before asking it to write automation programs or reconfigure critical devices.
Understanding memory and processing speed
A model's parameter count (such as 4B or 9B) is only one factor in overall performance. Total memory usage also depends on quantization level, context window size, runtime buffers, and other programs running on your computer. The size of the GGUF file on disk is not the total RAM required at runtime.
Lailama can run purely on your CPU, or use a supported GPU backend to accelerate inference:
- GPU Acceleration: A dedicated graphics card with sufficient VRAM can dramatically increase response speed.
- Integrated Graphics: An integrated GPU shares system RAM with your processor. Under sustained inference, it can generate noticeable heat and trigger system cooling fans.
- CPU Inference: Reliable on any modern computer, but complex reasoning and large coding prompts may take a minute or two to complete.
Always leave sufficient memory for your operating system, core HomeGenie automations, and any speech or camera services. If memory feels constrained, choose a smaller quantization level, reduce the context window, or switch to a lighter model.
You control privacy and permissions
Lailama maintains a strict separation between the data it can see and the actions it can perform. The toolbar above the chat provides separate Context and Capabilities menus, stored independently for each Genie.
| Context | Information available to the assistant |
|---|---|
| Devices | Module names, addresses, types, and real-time status, with feature details available on demand. |
| Automation Programs | Descriptions of your programs and their source code when relevant to the task. |
| Widgets | Available dashboard widgets along with their HTML, CSS, and JavaScript source code. |
The setting Share system and module context with the AI acts as a global master switch. Keep it enabled whenever you want the assistant to inspect your devices and programs, and then use the chat menu to choose exactly which categories to share.
| Capability | Actions it enables |
|---|---|
| Device control | Controlling devices and applying supported configuration changes. |
| Automation | Creating, updating, and running automation programs. |
| Widgets | Generating and editing custom dashboard widgets. |
For example, enabling the Devices context while disabling Device control allows Lailama to check temperatures and light levels without having permission to switch devices on or off. HomeGenie validates permissions before executing any command; the model cannot grant permissions to itself. Background requests follow these exact same rules.
Transparent actions and verification
When you ask about a configurable feature, Lailama can look up its schema, find the program that manages it, and read the latest parameter values through HomeGenie's APIs. The system retrieves fresh values during the conversation rather than relying on cached assumptions.
Whenever Lailama performs an action or reads a setting, the command and its raw response appear in an expandable API command block. This transparency allows you to clearly separate real system data from AI-generated text. Saved chat logs are purely historical: reopening an earlier conversation will never re-execute past commands.
Local models can occasionally misinterpret complex phrasing. If a response seems unexpected, expand the API block to verify what happened behind the scenes. Using clear device names usually produces immediate results; providing an explicit module address can help if you are experimenting with very small models.
Create programs and widgets locally
With the Automation capability enabled, Lailama can generate complete HomeGenie programs in C#, JavaScript, or Python. You can create brand new programs or ask the assistant to modify an existing one while keeping its program ID intact. If a new program reports a compilation error during installation, HomeGenie automatically feeds the error details back to the model so it can fix the code.
With the Widgets capability enabled, the assistant can write custom user interfaces with HTML, CSS, and JavaScript, displaying an interactive live preview inside the chat. When you are pleased with the design, click Save to add it to your dashboard. You can also fine-tune the result in the Widget Editor using Update, Save As, or Auto-update.
You can even ask for full-stack solutions in one prompt: a background program paired with a custom widget that displays its output. Because coding and UI generation require higher reasoning skills, pairing a capable local model with clear prompts delivers the best results. You can also use Gemini Genie for heavy coding tasks while keeping Lailama dedicated to private, everyday home control.
When a program finishes a one-time routine, its status displays as Completed, leaving it installed and ready to run again. A Quit command stops and disables the program while keeping its source code safely preserved.
Fine-tuning your conversation settings
You can easily adjust Lailama's runtime behavior to match your hardware and preferences:
- Creativity (Temperature): Sets how creative or deterministic responses will be. Lower values produce more focused, predictable answers.
- AI Memory Capacity (Context Window): Configures the working memory budget for the conversation.
- Response Length (Max Tokens): Sets the maximum length of generated replies. (Reasoning models use part of this allowance to think before answering).
- Chat Memory (Max Turns): Limits how many conversation turns are sent with each prompt, keeping processing fast without deleting your visible chat history.
- System Prompt (Instructions): Lets you personalize the assistant's personality and rules alongside HomeGenie's built-in capability instructions.
- AI Model List: Gives you full control over your model library, including download links, custom templates, and hardware notes.
Long chats and large source files naturally require more context and computation time. If a model seems slow, starting a fresh chat session will immediately restore peak responsiveness. Clearing the chat history only resets the working memory; it never requires reloading the model weights.
While a request is being processed, you can click Stop at any moment to cancel it safely. Canceling an in-flight prompt will not undo device actions that have already finished, and interrupted text will be marked as [STOPPED].
Local voice control with Whisper and Piper
Lailama works seamlessly with HomeGenie's offline speech tools to give you a fully voice-enabled smart home without third-party cloud services:
- Whisper provides accurate local speech-to-text directly from your microphone.
- Piper delivers fast, natural-sounding voice replies locally.
Both tools run on your HomeGenie server, while audio capture and playback happen smoothly inside your browser. Once enabled in your chat options, you can speak your requests naturally and even review or edit the transcription before sending it. (Note: browser microphone access requires a secure HTTPS connection or a local localhost address).
For detailed installation instructions, visit the Setup guide. Speech models operate independently from your language model and use their own separate memory.
Background automation and scheduled routines
Your automations and custom scripts can use Lailama without opening the interactive chat:
Request.Generate: Generates text summaries, status messages, or custom logs without triggering device actions.Request.Execute: Performs multi-step tasks through HomeGenie's action system within Lailama's permitted context and capabilities.Request.Status: Retrieves the current execution progress and result of an active request.Request.Cancel: Cancels an active request gracefully.
You can access these features through the shared System.Utility/AiGenie/ service or directly under AI.Genies/Lailama/.
In the Scheduler, you can add a Genie Command action, select Lailama (or Default), and write your task in natural language—for example: “Check that all exterior doors are locked and turn off the garden lights.” The scheduler pauses until the assistant finishes before moving to the next task, ensuring reliable automated routines.
You can run Lailama and Gemini Genie side by side on the same system. Each assistant maintains its own independent conversation history, context choices, and capabilities, giving you complete freedom to use the ideal AI for every task.