Privacy AI: Powerful chatbotv2.3.0

What changed

• New featureNew Email Assistant mini app for triaging, summarizing, and organizing IMAP inboxes.
• UXAbility to assign custom names to individual steps in AI Keyboard pipelines.
• New featurePersonal library to save and reuse named pipeline steps across different workflows.
• New featureMLX vision models now support video input with configurable frame counts and rates.
• PerformanceTurboQuant compression for long chats to reduce on-device memory usage.
• New featureDirect import of MLX models from local ZIP files without requiring downloads.
• PerformanceLocal engine update enables GPU support for two-bit models on Apple Silicon.
• PerformanceSignificant memory savings for vision models and improved multi-turn image chat consistency.
• UXEngine-aware settings that only display options relevant to the active model's engine.
• PerformanceFixed issues with memory profile display, model duplication, and context window settings.

Original release notes

FEATURES
- Email Assistant
A new mini app that triages your inbox. Connect any IMAP account with an app password, tap Process Inbox, and the AI reads unread mail, flags what needs a reply, and files the rest into folders under an AI label. Choose to summarize, organize, or also move obvious junk. Each run makes a dated report where every decision undoes with one tap, and your corrections teach your preferences. Passwords stay in the Keychain, content is never stored, and nothing is deleted. Mail can also become calendar events, reminders, and notes.
- Name Your Pipeline Steps
Give each step in an AI Keyboard pipeline its own name, so it shows your label instead of "Custom".
- Reusable Saved Steps
Save any named step to a personal library and reuse it across pipelines, each an independent copy.
- Video on Local Vision Models
MLX vision models can now understand video, not just images. Attach a short clip; set the frame count, rate, and size per model.
- TurboQuant for Long Chats
An opt-in option that keeps your whole conversation in on-device memory by compressing cached values, so long chats use less memory with a small quality trade-off.
- Import MLX Models from ZIP
Add your own MLX model from a ZIP on your device, no download needed. Tap Import From Local, pick a zip with the model folder, and it installs.

IMPROVEMENTS
- Local Engine Update (llama.cpp b10091)
Two-bit (Q2_0) models now run on the GPU on Apple Silicon, and Qwen3-VL reads layouts more accurately. Adds Laguna code models and completes DeepSeek V4 support.
- MLX Engine Update
Major memory savings for vision models: Qwen3.5 VL and Qwen3 VL handle long and multi-image prompts with far less memory, and multi-turn image chats stay consistent. Generation also stops cleanly when the app backgrounds.
- Engine-Aware Model Settings
Settings now show only the options the model's engine actually uses, each tagged with its engine. MLX models now respect Top-K and Min-P.

BUG FIXES
- Memory Profile Stays On Topic
The memory profile now appears only when memory is on and a sector is selected, and no longer blends unrelated sectors together.
- Dismiss The Suggestion Bar
The word suggestion bar can now be closed with an X, and no longer covers the pipeline picker.
- No More Duplicate Imported Models
Fixed a model imported from a local GGUF file reappearing as extra "Recovered from" copies after relaunch. Earl…