[macOS Apple Silicon][MLX] Qwen3.8 27B 6-bit is scanner-parsed and backend-supported but omitted from catalog / lms ls

[SantiagoMorenoGonzalez]

Expected

After lms get --mlx successfully downloads a supported MLX model, the model should appear in lms ls and be resolvable by lms load.

Actual

lmstudio-community/Qwen3.8-27B-MLX-6bit downloads successfully and is parsed by scanner/cache metadata, but is omitted from the active catalog and cannot be resolved by the CLI.

Environment

• LM Studio: 0.4.20+1
• lms CLI revision: 71bd99c
• Platform: macOS on Apple Silicon
• Model: lmstudio-community/Qwen3.8-27B-MLX-6bit
• Architecture: qwen3_5
• Quantization: MLX affine 6-bit

Reproduction

1. Download with the supported command:

   lms get --mlx --yes https://huggingface.co/lmstudio-community/Qwen3.8-27B-MLX-6bit

2. Confirm the downloaded artifact is complete. I independently compared it with a preserved previously working copy: the copies were byte-identical across 15 files (config, model.safetensors.index.json, tokenizer/config metadata, and all five model shards), totaling 22,804,836,386 bytes.

3. Run lms ls. The Q6 model is not exposed as a loadable catalog entry, while the generic 5-bit variant remains visible.

4. Attempt supported load/estimate resolution using both the model directory and qwen/qwen3.8-27b@6bit. Both are reported as unknown model keys and fall back to the interactive generic-model picker. No Q5 model was selected or inferred.

Evidence

• Scanner/cache metadata identifies the Q6 as qwen3_5, 6-bit affine, with exact local-model and virtual @6bit metadata.
• Historical LM Studio logs show this same Q6 artifact was previously loaded successfully, with fitted context 80896.

This suggests:

• backend support: yes
• artifact validity: yes
• scanner parse support: yes
• active catalog/resolver admission: no

Current classification: LMSTUDIO_CATALOG_STATE_BUG.

Questions

1. Is there a supported command or action to rebuild, rescan, refresh, or re-register LM Studio's local MLX model catalog?
2. Why can scanner/cache metadata parse this model while lms ls and lms load do not admit it?
3. Is dense qwen3_5 MLX 6-bit Qwen3.8 expected to be catalog-loadable in LM Studio 0.4.20+1?
4. Has this exact catalog/resolver issue been fixed in a later version/build? If so, what exact LM Studio version, build, and relevant issue or commit?
5. Is there an officially supported way to load this already-scanned local MLX directory without relying on the broken catalog entry?

I have intentionally omitted absolute local paths, usernames, credentials, session data, and unrelated application details.