top of page

Layla v4.0.0 has been published!

Welcome to the next major version of Layla!

First of all, thank you everyone for the support; it is the only reason Layla has evolved thus far!

Layla v4 brings significant changes to both features and UI. In this article, we will highlight some of the major changes that have been released.

Base Model Changes

Layla v4 uses completely re-trained base models. Initial benchmarks and real-world testing ( show very promising results.

Layla Lite is also re-trained using the same dataset. Initial benchmarks here shows that it outperforms the current Layla Full at all levels. Full characters load 10x faster using the new Lite model compared to the current Q2 and Full with same level of intelligence.

From now on Layla Lite is always the recommended model. I have been testing/developing/using Layla almost exclusively with the Lite model, and the performance is reasonable for most tasks. Yes there are a few hallucinations here and there, but the increase in speed is worth it.

Context extension

Implementation of Self-Extend ( allows us to extend the context of our current models 4-16x without any fine-tuning. This will be added as a configurable option in Layla. Very good for summarising long articles, long-form roleplay, and of course...

Long term memory

Conversation -> knowledge graph -> embeddings -> similarity search -> RAG gives us a good approximation of long-term memory by injecting relevant pieces of information during conversation. Additionally, combining this with putting dynamic information in the prompt allows characters to recall past information between conversations. This is made possible by both the increased loading speed of the Lite model and the context length extension mentioned above.

For a more in-depth article on how it works, see:

New tools in Layla's toolbox

  • Web search updated - parses your search intent better with keyword detection to give you more relevant results.

  • News summary - split this feature from the web search so when you request for current news/events, the results come back focused on news articles instead of general web search to reduce hallucinations.

  • To-do list - has been re-integrated as part of Layla's toolbox to allow you to schedule reminders and calendar events during conversation.

UI updates

  • Conversation history page has been redesigned to allow you to scroll through histories of different characters easier, with ability to delete individual chats.

  • You can configure Layla to start from the "Messages" page if you like an experience that's similar to messaging apps such as Facebook Messenger, Snapchat, etc. with your characters.

We hope you enjoy this update, and as always, we welcome any feedback to better improve Layla!

114 views0 comments

Recent Posts

See All


Rated 0 out of 5 stars.
No ratings yet

Add a rating
bottom of page