Expanded the experimental text catalog with MiniCPM-V 4.7, LFM2.5, Jais 3 and Nemotron 3.5 Bolt, plus complete multipart Beam and Muse installation. Added separate workflows for one-image questions, approved Hy-Memory proposals, verified Granite Switch 4.1 imports and local image generation on compatible high-memory hardware. Added an optional Step 5 connection to your own prepared CUDA host, with approval before sending its conversation. Improved compatibility, resource and complete-file checks. Clarified local and optional remote processing in the app and privacy policy.

Added experimental Bonsai 2 2B, 4B and 9B choices, with model-specific memory checks and pinned downloads. Added experimental native text support for Muse Spark Open, including complete multipart verification, resumable installation and final-answer filtering. Muse requires a high-memory machine and is unavailable on ordinary phones. It is never selected automatically. Improved model compatibility checks and recent-release discovery. Your model selection and conversations remain under your control.

• Added experimental Ternary Bonsai 2 27B to the model catalog, with the Prism runtime included. Select it manually on devices with enough memory; the download is about 6 GB and running it needs additional RAM. • Added model release discovery and download-size filters. • Improved cancellation and handling of long prompts. Chat generation stays on your device. The default model recommendations and memory checks remain in place.