New Chat
No model selected
OFFLINE
Your chats and settings now follow your HOME LLM account across devices.
Make the assistant feel more like you. These preferences change its style, not its capabilities.
The overall voice of your assistant.
How reassuring and personable responses feel.
Adjust the energy in responses.
How often responses use headings and lists.
How often the assistant uses emoji.
Optional name used in replies.
School, work, or interests that help tailor examples.
Background, goals, and preferences you want included.
Additional behavior, style, and tone preferences. Sent to your home model with each request.
Prefer a quick answer or a more detailed explanation.
Choose how much background the assistant should assume.
Auto follows the language of your message.
Adjust how often concrete examples are included.
Choose when the assistant should ask before answering.
Ask the model to distinguish guesses from facts.
Used when you have not specified a language.
Adjust the amount of explanation inside generated code.
Preferred units when your request does not specify them.
Applied to future requests. Larger token limits and thinking levels can increase wait time and memory use.
Higher values produce more variety. Blank keeps the model default.
Limits the pool of likely next tokens. Blank keeps the model default.
Values above 1 discourage repetition. Blank keeps the model default.
A fixed seed makes runs more repeatable, but not guaranteed identical. Blank is random.
Maximum generated tokens per normal answer. Fast Mode uses a smaller budget.
Larger windows fit more conversation and use more memory. Limited by the model.
Normal answers only. GPT-OSS supports low/medium/high; other thinking models use on/off. Fast Mode and internal reviews keep low/off thinking.
Real separate model calls. Image reviews require a vision model.
Also controlled by the lightbulb in the composer.
Prefer concise answers with a smaller output budget.
Stop when the current draft reaches this reviewer score.
Maximum review/rewrite rounds. Server limits can lower this value.
Flat dark layout inspired by the reference UI: compact sidebar, plain assistant responses, dark user bubbles, pill composer, and matching message menus.
Original black background, white outlines, and green connection indicator. ChatGPT mode and OG mode cannot be enabled together.
Choose light, dark, or follow your device. ChatGPT and OG modes use their own dark palettes.
Color for active controls, buttons, and highlights.
Adjust message readability.
Space between messages.
Maximum width of the conversation.
Used for messages; code remains monospace.
Display token counts and generation speed.
Scroll with new text when you are already near the bottom.
When off, use Ctrl+Enter or Cmd+Enter to send.
Wrap long lines instead of scrolling horizontally.
Turn off interface transitions and animated effects.
When HOME LLM is opened from its HTTPS server address, Chrome/ChromeOS can install it as a standalone app. The portable HTML still works separately.
Local suggestions in empty chats. No plugin searches or extra model calls.
When off, current chats stay in memory only. Saving this change removes stored chat history.
Data actions happen immediately. Exports contain chat text and attachment references, not uploaded file contents, login tokens, or admin settings. Imported attachments may require uploading again on another browser.
Your LAN address or HTTPS tunnel URL.
Chats, projects, presets, and settings sync to your account. Server-wide admin controls remain under the HOME LLM logo.
No matching settings.
Manage your HOME LLM server, hardware, users, limits, and assistant behavior.
Refreshes while Administration is open.
Sent as part of one server-owned system message. The same saved version is used throughout a request. Natural-language instructions still depend on the model.
Prefix and suffix are added by the server to completed answers, outside the model. For a guaranteed “meow” ending, put a space followed by meow in the suffix. These additions are not included in model-token counts and may affect code-only answers. Use these controls instead of asking the model to add the same text.
Used when a browser has no valid selection. Existing selections stay unchanged.
Changing the code signs out existing users.
Recent jobs show operational metadata only. Chat prompts and responses are not logged here. History clears when the server restarts.
Stop finishes the current model call before releasing the GPU, then skips remaining work.