Lekh AI - Local AI Studio

- A new Large MoE Models section in the catalogue has models with up to 35B parameters that run on your device: Gemma 4 26B, Qwen 3.6 35B, Qwen 3 30B, Qwen 3 Coder 30B, GLM 4.7 Flash, Nemotron 3 Nano 30B and Ling Mini 2.0. - They keep memory use low by loading only the parts they need from storage as they go. - Each chat starts with a clean slate, with no system prompt. Faster on-device AI - Gemma 4 runs several times faster, and thinking now works for it. - The chat stays smooth while a model is writing, including when you scroll. Better thinking - A new thinking box shows the model's reasoning as it happens, then folds away when the answer starts. - Qwen 3.5 and other reasoning models now put their thinking in the box instead of in the answer. Better formatting - Replies now show headings, bullet and numbered lists, quotes and dividers. - Long replies keep their formatting instead of turning into raw text. Fresh look - A new soft gradient behind the chat and the sidebar. - Cleaner, bolder sidebar and toolbar buttons. Additional Support - the Apple Intelligence model (iOS 26 and later) - Bonsai MLX models - the thinking on/off button - the iPhone Duo layout with its Tools tab Fixes - Fixed a crash when leaving the app while a reply was being written. - The model picker no longer disappears when you start a new chat. - Fixed a model that kept writing after it had finished its answer.