Layla Update v6.12.0: Roleplay Mini-App Redesigned With 50+ Templates, Local Model Support, and Multi-LLM Conversations

Layla v6.12.0 has been published!

Layla's Roleplay mini-app has been rebuilt from the ground up. This update adds a library of 50+ ready-made templates, support for running roleplay on local GGUF models as well as remote backends, automatic switching between LLMs depending on who is speaking, a Visual Novel display mode, out-of-character instructions, per-character text-to-speech, and a new editor for building your own scenarios.

A redesigned Roleplay mini-app with 50+ templates

The fastest way to start is now a template. There are more than 50 to pick from, covering romance, adventure, isekai, comedy, interview practice, language learning, and others. Each one drops you into a configured scenario so you can start a session without building everything from scratch first.

Templates are a starting point — once a session is running you can edit the scenario, swap characters, or branch off in whatever direction you want.

You can add participants, generate images, tag your scenarios and more right in the creation screen.

Roleplay now runs on local models, or any backend you choose

Roleplay is no longer tied to a local LLM. You can run it on:

• Local GGUF models, fully on-device

• A connection to your PC or an OpenAI-compatible endpoint

• Layla Server

• Layla Cloud

If you want a roleplay session that never leaves your phone, point it at a local GGUF model and it runs entirely offline. If you'd rather use a larger model on your desktop or a hosted endpoint, the same mini-app connects to those instead.

Multiple LLMs in a single roleplay

Roleplay now switches LLMs automatically based on your saved inference engines, choosing the right one depending on who is talking. The practical result is that you can have multiple LLMs in the same conversation and watch how they interact with each other, with each character driven by whichever model you assigned to it.

This pairs with the redesigned Saved Inference Engine section (see Improvements below), which makes it easier to see which engines are attached to which characters.

The new scenario and character editor

There's an all-new editor for building or modifying scenarios. You can:

• Create temporary characters directly inside the Roleplay mini-app, without leaving to set them up elsewhere first

• Attach your existing characters to a scenario

Temporary characters you create this way are filed automatically into a folder named after the roleplay, so they stay grouped in your character list instead of cluttering it.

Visual Novel mode, OOC instructions, and per-character voices

Three additions change how a session reads and plays out:

• Visual Novel mode centres the character's image as they speak or act, for a more visual-novel-style presentation.

• OOC (out-of-character) instructions can be added at any point during a session. They're useful for steering the AI back on track without breaking the scene or editing the scenario.

• Text-to-speech per character gives each character its own voice. Voices are configured in the character edit screen.

Roleplay also supports custom agents and image generation, so a scenario can run tools or generate images as part of the session.

Improvements

• Improved long-term memory summarisation speed.

• Added a Gemma 4 (thinking) preset prompt.

• Redesigned the Saved Inference Engine section to make it easier to filter by attached characters.

• Improved TTS consistency by removing unnecessary mid-sentence stops.

• Updated Local Dream with image editing features.

• Added new schedulers (Karras versions) in Local Dream.

Bug fixes

• Fixed a bug where the model selector disappeared after searching for available models when configuring the OpenAI API.

• Fixed a bug where chat bubble styles stopped showing when the performance counter was enabled.

• Fixed a bug where changing the prompt but not the seed in image generation did not produce a new image.

• Fixed a bug where "Speak Responses" in chat actions could not be toggled off.

• Fixed a bug where "Push to talk" had stopped working.

When you run Roleplay on a local GGUF model, the whole thing stays on-device — no account, no cloud, and no data leaving your phone.