Pick your model. Talkink now ships a catalog of seven on-device speech models — choose the one that fits your Mac and your languages, or grab several and switch instantly.

✨ New

• Model catalog in Settings — Qwen3-ASR 1.7B & 0.6B, NVIDIA Nemotron 3.5 and Voxtral Mini 4B (8-bit / bf16 / 4-bit). Real download sizes, quality & speed ratings from our own multilingual benchmark, and a clear "On this Mac" status per model. Recommended default: Qwen3-ASR 1.7B 8-bit — the most accurate model we measured.
• Resumable downloads — live byte-accurate progress bar, several models downloading at once, and an interrupted download (quit, crash, Wi-Fi drop) picks up exactly where it stopped: "Resume — 62% here".
• Delete models you no longer use, right from the picker — every gigabyte is yours.
• 27 dictation languages (up from 9) — each verified against what the models actually support, and passed to every engine in its own native format.
• Wrong-language rescue — picked Swedish but spoke French? When a model returns nothing for good speech, Talkink silently retries in auto-detect so your dictation lands anyway.
• Double-clicking Talkink in Applications now brings up its window, like any Mac app.

🔧 Improved

• Switching models frees the old one from RAM (measured: leaving the 1.7B model releases ~1.6 GB). One model in memory, never a stack.
• The footer and About panel name the model you're actually using.

⬇️ Install

Download Talkink.dmg, open it, drag Talkink into Applications. Notarized by Apple — no warnings, no terminal.

Existing users: Talkink will offer this update automatically.