0.16.10 — Easier downloads and immediate dictation

On restricted networks, covered catalog models now download from the Rapid-MLX mirror without contacting Hugging Face first. Desktop dictation starts capturing as soon as you invoke it, so the beginning of a sentence is kept.

• Start a model with fewer network detours. Covered, pinned model files cold-pull from the mirror and still verify their expected contents, including files in subfolders. A cancelled pull now shows a retry command. Models outside mirror coverage may still need Hugging Face. (#5293, #5248, #5283)
• Dictate from the first word. Desktop begins audio capture immediately. Complete, pinned speech-to-text models and cached Whisper processor files can also load when Hugging Face is unavailable; missing files still require a connection. (#5275, #5301)
• Choose from three more models. Nanbeige4.3-33B, Spark-X2.6-4B and Sage-A1.5 have pinned MLX aliases. LFM2 MoE single-token decoding is faster on the tested Apple Silicon systems; the reported gain is specific to that workload. (#5279, #5298)
• Keep conversations on track. Cache reuse prefers longer prefixes that can actually be restored, with a fallback if restoration fails. Output handling improves reasoning and tool markup in supported paths while preserving streaming behavior. (#5253, #5264, #5250, #5296)