SayFlow: AI Chat, Translate

[v2.0.1] What's New • Added a Local Model provider that runs AI models entirely on the device, with no API key and no data sent to a server. • Merged Apple Intelligence into Local Model. Choose Apple Intelligence, Gemma 4 E2B, Qwen3.5 0.8B, Qwen3.5 2B, or your own GGUF model. • Apple Intelligence is the default on supported devices. Gemma 4 E2B is the default on other devices. • Models are offered based on the device's available memory (RAM) instead of the device model. • Added a Manage Models screen: - Download a model with the download icon, then watch its progress and cancel if needed. - Delete a downloaded model with the trash icon, after a confirmation prompt. - Each model shows a one-line description, its size, and its memory status. • Downloads continue in the background, even if the app is closed. • Import your own GGUF models from the Files app. SayFlow checks the file format, chat format, model architecture, and memory before importing. • Added image input for Gemma 4. After downloading Image Input (557 MB), you can attach photos to local chats. • Local models are available in Settings and in the chat model menu, with model logos. Each chat remembers its model. • Updated the provider icon and related settings text. Chat • Sent messages now rise from the composer into place. • The keyboard now stays open after sending a message. • Bold and italic text now render correctly next to inline math. • The "Writing a reply…" bar no longer covers the composer while a chat or translation screen is open. Fixes • Fixed the control widget label. • Fixed an issue where a chat could suddenly turn blank right after opening it. • Scrolling in conversations is steadier, with fewer unexpected jumps. • Sent messages now glide into place with a smoother animation. • Tap anywhere in the conversation to dismiss the keyboard. • Automatic chat compression now waits until a conversation is actually long, so short chats are no longer summarized after just a few messages. • Compressed summaries now build on the previous summary, making repeat compressions faster and more accurate. • The keyboard now stays open after sending, so you can keep typing.