Chat LLM

This update adds a full document reader, layers spoken playback over every AI reply, slims the app's install footprint dramatically, and rolls in the major voice-chat overhaul plus web-search feature from the previous release.

Reader - listen to anything
A full audiobook reader is now built in. Import a PDF, EPUB, or TXT file, or paste any text, and Reader speaks it sentence by sentence using Kokoro voices via Replicate. Tap any sentence to jump to it, pick your voice and speed, and your position is saved automatically. Accessible from the model picker alongside Chat and Voice. Requires your own Replicate API key (~$0.00088 per request).

Spoken playback of any AI reply
Tap the play button next to any AI message to hear it read out loud. 30+ Kokoro voices to choose from, adjustable playback speed, full lock-screen and Control Centre controls, and an audio cache so the same message plays instantly the second time. Bring your own Replicate API key and generation costs go to your own Replicate credits, roughly $0.00088 per reply.

Smaller install, on-demand local model
The app is now under 10 MB to install, down from over 700 MB. The on-device Llama 3.2 model has moved to an optional one-tap download from the model picker, so you only fetch it if you actually want offline AI. Existing local-model users: tap the Local LLM card to bring it back in a couple of minutes.

Voice chat now powered by Google Gemini Live
We've moved voice chat from xAI to Google's Gemini Live for a faster, more natural realtime conversation. Many new voices to choose from, and your Gemini API key now powers it (get yours free from Google AI Studio).

Show, don't tell - image input in voice chat
You can now share an image during a voice conversation. Hold up a photo or screenshot and ask the assistant about it while talking. Powered by Gemini Live's multimodal input.

Deeply customisable voice personality
The voice assistant's personality, accent, speech style, vocabulary, language preferences, and behaviour can all be shaped through the system prompt. Set the kind of person it is, how it speaks, what units it uses, what topics it leans into, make it your own.

Web search for OpenRouter chat (BYOK)
A "Web Search" toggle in the menu lets any OpenRouter model pull live information into its answers — current events, news, prices, weather, anything time-sensitive. Toggle on, ask, get an up-to-date reply. This release also adds device-locale awareness, so users outside the US get region-relevant results on the first attempt (UK searches return UK sources, and so on). Charged to your own OpenRouter credits at OpenRouter's standard per-result rate.

Cleaner model picker
Capability tags (Vision, Reasoning, parameter count, etc.) now appear correctly on every OpenRouter model, including the "always latest" alias variants that previously showed up bare.

Layout polish
Bigger thinking box for reasoning models - easier to follow long chains of thought.
Wider message bubbles in landscape on iPhone and iPad.

Terms of Use (EULA): https://www.apple.com/legal/internet-services/itunes/dev/stdeula/

https://ferrraridave-coder.github.io/chat-llm-support/privacy-policy.html