The Layla Chat Screen Explained: Every Control, Menu, and Gesture

Layla's chat screen is a single scrolling conversation with a fixed input bar at the bottom and a header at the top that tells you which character and which model you are currently talking to. Everything else on the screen is reachable from one of those two strips.

This guide walks through the screen in the order you actually encounter it: what you see before you send anything, what changes while a response is being generated, and what you can do with a message once it has finished.

Before you send: the resting state

When you open a chat, the screen has three zones.

The header. This shows the character you are talking to and gives you access to chat-level menus. The character name and avatar here are not decoration — they determine the system prompt, the voice used for text-to-speech, and any Lorebook attached to that character. If responses are coming out in a personality you did not expect, the header is the first place to look.

The conversation area. Your messages and the character's responses, oldest at the top. Scroll up to read history; the conversation is stored locally and is not truncated on screen even when older turns have fallen out of the model's context window. That distinction matters and confuses people regularly: seeing a message in the scrollback does not mean the model can still see it. Context length is set in advanced settings, and older turns beyond it are handled by long-term memory summarisation rather than being sent verbatim.

The input bar. The message field plus its surrounding controls. In its resting state the input bar shows:

• The message text field, with a lightbulb to access your placeholder prompts.

• A send button.

• A wave control icon for voice input.

• An attachment or plus control that opens the secondary actions.

The message field grows as you type past one line and scrolls internally past a certain height, so long prompts do not push the conversation off screen.

The model and character indicator

The character is the personality: the system prompt, greeting, example dialogue, voice, and any attached Lorebook. The model is the actual LLM doing the generation — a GGUF model via llama.cpp, a LiteRT-LM model, or a remote API if you have configured one.

These are independent. Swapping the character does not swap the model, and loading a different model does not change who you are talking to. A character can also have an inference engine attached to it, so that selecting that character loads its preferred model automatically. If you have set that up, changing character will change the model — which is usually the explanation when someone reports that generation speed changed for no apparent reason.

You can see the debug info for a chat which shows the model and system prompt, which you can use to debug any unexpected responses. You can open this in the top right menu.

Sending a message

Type into the field and tap send. From that moment the screen changes state.

Prompt processing. Before any text appears, the model has to ingest the conversation so far. On the first message of a session this is the longest wait, because the entire prompt — system prompt, character definition, lorebook entries, history — is being processed. Subsequent messages are faster, because most of that work is cached. This is why the first reply after opening a chat feels slow and the rest do not, and it is the single most common "is it frozen?" report.

Token streaming. Text appears progressively, a few tokens at a time, rather than arriving all at once. If you have the performance counter enabled in settings, you will see tokens-per-second alongside the response while this happens.

The send button becomes a stop button. While generation is running you can halt it. The partial response is kept in the conversation rather than discarded, so stopping a reply that has gone off the rails leaves you something to edit or regenerate from.

What you cannot do mid-generation. You cannot send another message before the character has finished responding or you have stopped generation.

Voice input

Layla supports speaking your message instead of typing it, and speaking responses back.

Push to talk captures voice input from the chat screen.

Speak Responses is a toggle in Chat Actions that reads generated replies aloud using the character's assigned voice. Over 100 voices are available, and TTS skips content it cannot pronounce sensibly, such as ASCII art and tables, rather than reading punctuation aloud.

Both run on-device. Nothing is sent to a speech service.

Chat Actions and the chat menu

Chat Actions is the per-conversation control panel. It holds things that change how the current chat behaves without changing the character permanently — including the Speak Responses toggle and the ability to give your character or model a custom instruction for the current conversation only. That last one is useful for one-off steering ("keep replies under three sentences") without editing the character definition.

Per-message controls

Long-pressing or tapping a message exposes actions on that individual turn.

Regenerate and alternate responses are worth calling out in the article body once confirmed, because they are the fix for the most common complaint — a reply that is nearly right — and many users never find them.

Loading states you will see

What you see	What is happening
Model loading indicator on entering a chat	The model is being read into memory. Larger models and higher quants take longer.
No text yet, stop button active	Prompt processing. Longest on the first message of a session.
Text appearing incrementally	Token generation. Normal.
Generation never starts	No model loaded, or insufficient free RAM for the selected model and context length.

FAQ

Why is the first response so much slower than the rest?

The full prompt — system prompt, character definition, Lorebook, and history — has to be processed before the first token can be produced. After that, most of it is cached.

Can I stop a response halfway?

Yes. The send button becomes a stop button during generation.

Does Layla need internet to chat?

No. Text generation, voice input, and text-to-speech all run on your device. Features that do require a connection are shown prominently and ask for consent first. (If you select Layla Cloud as your inference setting, then it requires internet.)

Why did my character start replying differently after I changed models?

The character controls personality; the model controls capability and style of execution. A smaller or more heavily quantised model will follow a complex character definition less closely.

Where do I change the model?

You can adjust your model in Inference Settings, which you can access either in your mini-apps or settings page.

The chat screen is deliberately shallow — one conversation, one input bar, and menus that stay out of the way until you need them. Everything on it runs on the device in your hand, with no account and no connection required.