fix(mlx): point Qwen 0.6B at the published mlx-community repo

[jamiepine]

Summary

• Point Qwen 0.6B at mlx-community/Qwen3-TTS-12Hz-0.6B-Base-bf16 instead of falling back to the 1.7B repo.
• Drop the obsolete "0.6B downloads the same as 1.7B" troubleshooting entry.
• Clean up the stale triage note in PROJECT_STATUS.md.

Why

When the MLX backend shipped, mlx-community hadn't converted the 0.6B base yet, so the 0.6B slot was aliased to the 1.7B repo as a temporary fallback (# 0.6B not available in MLX, falls back). That conversion is now live (1.4k downloads, published 2026-01-22), so Apple Silicon users picking 0.6B should get the actual 0.6B weights — ~1.2 GB instead of ~3.5 GB — with the corresponding speed/memory win.

Reported in #485, also covers the older #423 / #329 reports.

Qwen CustomVoice is unaffected — it runs via PyTorch on all platforms and both sizes already have dedicated repos.

Test plan

• On Apple Silicon: delete any cached 0.6B model, select "Qwen TTS 0.6B" in the app, confirm the downloaded repo ID is mlx-community/Qwen3-TTS-12Hz-0.6B-Base-bf16 (not the 1.7B one) and the download size is ~1.2 GB.
• Generation works end-to-end with the 0.6B MLX model.
• 1.7B selection still resolves to the 1.7B MLX repo.
• On PyTorch (Windows/Linux): no behavior change — 0.6B and 1.7B already pointed at distinct Qwen/... repos.

Fixes #485.

🤖 Generated with Claude Code

---

Note

Medium Risk
Changes which HuggingFace repo is downloaded/loaded for Qwen 0.6B on the MLX backend, which can affect cache reuse and download behavior for existing users. Logic is small and localized but touches core model resolution paths.

Overview
Fixes Apple Silicon (MLX) model resolution so selecting Qwen TTS 0.6B downloads/loads mlx-community/Qwen3-TTS-12Hz-0.6B-Base-bf16 instead of incorrectly aliasing to the 1.7B repo.

Updates the MLX backend’s model-size mapping and the backend-aware model registry accordingly, adjusts the Qwen download test to remove the stale fallback note, and cleans up docs by removing/rewriting troubleshooting and project-status guidance that assumed the fallback behavior.

^{Reviewed by Cursor Bugbot for commit c70ec32. Configure here.}

Summary by CodeRabbit

• Bug Fixes

  • Fixed Qwen 0.6B model to correctly download dedicated 0.6B weights instead of falling back to 1.7B model on all platforms.

• Documentation

  • Updated troubleshooting guide to reflect resolution of the 0.6B/1.7B model mapping issue.
  • Removed obsolete workarounds and explanations related to the incorrect model fallback.

[coderabbitai]

No actionable comments were generated in the recent review. 🎉

ℹ️ Recent review info

⚙️ Run configuration

Configuration used: defaults

Review profile: CHILL

Plan: Pro

Run ID: fa5d48fe-cd34-4c27-b8b6-bd53e9b081c5

📥 Commits

Reviewing files that changed from the base of the PR and between e3f7cd9 and c70ec32.

📒 Files selected for processing (5)

• backend/backends/__init__.py
• backend/backends/mlx_backend.py
• backend/tests/test_qwen_download.py
• docs/PROJECT_STATUS.md
• docs/content/docs/overview/troubleshooting.mdx

💤 Files with no reviewable changes (1)

• docs/content/docs/overview/troubleshooting.mdx

---

📝 Walkthrough

Walkthrough

This pull request corrects a bug where selecting Qwen TTS 0.6B model incorrectly downloaded the 1.7B variant. The fix updates model repository IDs in the MLX backend configuration and removes stale fallback logic, with accompanying documentation and test updates.

Changes

Cohort / File(s)	Summary
MLX Backend Model ID Corrections backend/backends/__init__.py, backend/backends/mlx_backend.py	Updated MLX-specific repo ID for 0.6B TTS model from the 1.7B fallback to the correct mlx-community/Qwen3-TTS-12Hz-0.6B-Base-bf16 and removed fallback comment/logic.
Test & Documentation Updates backend/tests/test_qwen_download.py, docs/PROJECT_STATUS.md, docs/content/docs/overview/troubleshooting.mdx	Removed inline comment about 0.6B mapping to 1.7B in test, updated PROJECT_STATUS to reflect the stale MLX fallback has been resolved, and removed the dedicated troubleshooting section about the 0.6B/1.7B download issue.

Estimated code review effort

🎯 2 (Simple) | ⏱️ ~10 minutes

Poem

🐰 A rabbit hops through model IDs with glee,
The 0.6B now downloads correctly, you see!
No more 1.7B sneaking into the fray,
MLX fallbacks are fixed—hooray, hooray! 🎉

🚥 Pre-merge checks | ✅ 4 | ❌ 1

❌ Failed checks (1 warning)

Check name	Status	Explanation	Resolution
Docstring Coverage	⚠️ Warning	Docstring coverage is 66.67% which is insufficient. The required threshold is 80.00%.	Write docstrings for the functions missing them to satisfy the coverage threshold.

✅ Passed checks (4 passed)

Check name	Status	Explanation
Description Check	✅ Passed	Check skipped - CodeRabbit’s high-level summary is enabled.
Title check	✅ Passed	The pull request title clearly and specifically describes the main change: pointing Qwen 0.6B to the published mlx-community repository instead of a fallback.
Linked Issues check	✅ Passed	The pull request successfully addresses issue #485 by pointing the Qwen 0.6B selection to the correct mlx-community/Qwen3-TTS-12Hz-0.6B-Base-bf16 repository, resolving the bug where users were downloading the 1.7B model instead.
Out of Scope Changes check	✅ Passed	All changes are directly related to fixing the Qwen 0.6B repository mapping issue: backend code updates, test cleanup, and documentation updates that reflect the fix. No unrelated modifications are present.

_{✏️ Tip: You can configure your own custom pre-merge checks in the settings.}

✨ Finishing Touches

📝 Generate docstrings

• Create stacked PR
• Commit on current branch

🧪 Generate unit tests (beta)

• Create PR with unit tests
• Commit unit tests in branch fix/mlx-qwen-0.6b-repo

---

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

• X
• Mastodon
• Reddit
• LinkedIn

_{Comment @coderabbitai help to get the list of available commands and usage tips.}