docs: audio smoke evidence + audio-tower note + family status

README.md

@@ -16,7 +16,7 @@ tags:
 
 MLX 8bit (affine, 8-bit, group size 64) quantization of [`Qwen/Qwen3-ASR-0.6B-hf`](https://huggingface.co/Qwen/Qwen3-ASR-0.6B-hf) at upstream revision `6aa69c382e2b426eee1f5870d4c95859a74b6445`, produced directly from the BF16 safetensors with `pipelines.mlx_direct_quantize` (majek repo). ~0.9 GiB on disk.
 
-> **Runtime caveat:** this is a weights + config pack published ahead of runtime availability; runtime support pending upstream mlx-lm support for the `qwen3_asr` audio-text architecture (the Hugging Face runtime requires transformers 5.13+ — `qwen3_asr` merged after the 5.12.1 release). No local inference smoke was performed on these artifacts, so the quality of this variant is not verified.
+> **Runtime caveat:** this is a weights + config pack published ahead of runtime availability; runtime support pending upstream mlx-lm support for the `qwen3_asr` audio-text architecture (the Hugging Face runtime requires transformers 5.13+ — `qwen3_asr` merged after the 5.12.1 release). This variant **passed** the local MLX audio-inference smoke gate — see *Audio inference smoke* below (run natively with the source repo's `pipelines/qwen3_asr_mlx` MLX harness).
 
 ## Why this variant
 
@@ -28,14 +28,35 @@ Affine 8-bit, group size 64 — the highest-fidelity tier, roughly half of BF16
 python -m pipelines.mlx_direct_quantize --model qwen3-asr-0.6b --base-dir /tmp/mlx-direct-release/qwen3-asr-0.6b/base --out-dir /tmp/mlx-direct-release/qwen3-asr-0.6b/8bit --bits 8 --mode affine --group-size 64
 ```
 
+## Audio tower
+
+The `model.audio_tower.*` windowed-attention audio encoder (297 tensors) and the `model.multi_modal_projector.*` projector (4 tensors) are passed through unquantized in BF16 — only the `model.language_model.*` qwen3 text decoder is quantized.
+
+## Audio inference smoke
+
+**PASS** — this variant passed the audio-inference smoke gate on 2026-07-03.
+
+Gate design: character error rate (CER) is measured against an **MLX-BF16 self-reference** baseline (the same MLX audio path run at BF16). Only the clips that are stable and in-language for both family sizes (`1_cantonese.wav`, `4_codeswitch_singlish.wav` — Cantonese and Singlish) decide the verdict; the other clips are transcribed and recorded as informational. Threshold: CER ≤ 0.05 on every gating clip.
+
+| clip | CER vs MLX-BF16 | gating | wall_s |
+| --- | --- | --- | --- |
+| `1_cantonese.wav` | 0.0 | yes | 0.157 |
+| `2_hokkien.wav` | 0.988701 | no | 0.289 |
+| `3_codeswitch_tamil.wav` | 0.0 | no | 0.141 |
+| `4_codeswitch_singlish.wav` | 0.0 | yes | 0.066 |
+
+- Decoder-only smoke (one-token forward, finite logits): **pass**
+- Total wall clock: 39.753 s
+- Smoke run 2026-07-03 on Apple Silicon via the source repo's `pipelines/qwen3_asr_mlx` package (driver: `pipelines/qwen3_asr_mlx_smoke.py`).
+
 ## Family
 
 All MLX variants of this model ship together:
 
-- [majentik/Qwen3-ASR-0.6B-MLX-4bit](https://huggingface.co/majentik/Qwen3-ASR-0.6B-MLX-4bit)
+- `majentik/Qwen3-ASR-0.6B-MLX-4bit` (private — pending requantization; failed the audio-inference smoke gate)
 - [majentik/Qwen3-ASR-0.6B-MLX-6bit](https://huggingface.co/majentik/Qwen3-ASR-0.6B-MLX-6bit)
 - `majentik/Qwen3-ASR-0.6B-MLX-8bit` (this repo)
-- [majentik/Qwen3-ASR-0.6B-MLX-MXFP4](https://huggingface.co/majentik/Qwen3-ASR-0.6B-MLX-MXFP4)
+- `majentik/Qwen3-ASR-0.6B-MLX-MXFP4` (private — pending requantization; failed the audio-inference smoke gate)
 
 ## Provenance