[daniellee-ux]

Seminarly can now run Qwen3-ASR 0.6B alongside Whisper and reuse compatible models already present on the Mac. Users select a transcription model in Settings; the app finds supported local caches before downloading missing files. Whisper remains the production default.

Changes

• Add an optional native Swift/MLX Qwen backend for Apple Silicon, using pinned 0.6B weights with a 4-bit decoder and 8-bit encoder (~542 MB). Reuse the existing capture, transcript storage, and speaker-attribution pipeline; defer model switches during recording and restore the working model if a switch fails.
• Discover compatible models in Seminarly, Swift Hub, and Hugging Face Hub caches. Verify pinned Qwen files by size and SHA-256, load complete installations in place, and reuse valid files from partial snapshots without copying large weights. External caches remain read-only; generated tokenizer files live in app-owned storage. No Python runtime or background model server is added.
• Correct Qwen frontend frame/token-length handling and Whisper automatic-language/prompt-cache options. Add synthetic reference fixtures and opt-in inference checks.
• Provide a separately identified Seminarly Qwen Test app and packaging script. Test build 18 uses isolated preferences/database/update behavior while sharing model weights. New recordings do not retain raw audio; an optional legacy field remains for compatibility with existing test databases.

Validation

• Optimized QwenTest XCTest suite: 536 cases, 0 failures, 1 skipped because its optional old-database fixture was not supplied.
• Real Mandarin inference passed through both engines, including Qwen from a read-only Hugging Face snapshot and Whisper with separate model/tokenizer caches. Qwen engine silence/timeline checks passed.
• Existing production database migration: both focused compatibility tests passed using an isolated SQLite backup, including preservation of sessions and transcripts.
• Release metadata prepared for v0.1.16 (build 17).
• Test build 18 packaged successfully; app signatures and DMG checksum verified. Installed app launched and displayed Ready on this Mac with the existing session still visible.

Current limits

Qwen remains experimental and requires Apple Silicon. Its segments use approximate audio bounds of up to eight seconds, so rapid speaker changes and overlapping speech still need evaluation; this does not replace word-level alignment. Automatic discovery supports the offered Whisper Core ML variants and the exact pinned Qwen artifact, not arbitrary PyTorch/GGUF models. Inference checks establish functionality, not meeting-accuracy benchmarks.

[chatgpt-codex-connector]

Codex Review Summary

This comment shows the latest Codex review activity on this pull request.

Review	Status	Commit	Review trigger
📝 Code Review	✅ Completed 2026-09-28T09:27:28.092211Z	05bd115	PR opened

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review" or "@codex security review".

Codex reacts with 👀 while any review is running, comments if it has suggestions, and reacts with 👍 once all reviews finish with no findings.