# Provenance

| | |
|---|---|
| Upstream | [Qwen/Qwen3.6-35B-A3B](https://huggingface.co/Qwen/Qwen3.6-35B-A3B) @ `995ad96eacd98c81ed38be0c5b274b04031597b0` |
| Converted | 2026-07-04 |
| Toolchain | mlx-vlm 0.6.3, mlx 0.31.2, mlx-lm 0.31.3, transformers 5.12.0 |
| Command | `python -m mlx_vlm convert --hf-path Qwen/Qwen3.6-35B-A3B --revision 995ad96eacd98c81ed38be0c5b274b04031597b0 -q --q-bits 5 --q-group-size 64 --q-mode affine` |

Contents: language tower quantized (5-bit affine, group size 64) +
**vision tower in BF16** (333 tensors, unquantized — mlx-vlm skips
multimodal modules by default). The upstream MTP head is not retained
(dropped by mlx-vlm sanitize; unused at inference).

Smoke (pre-publish, Apple Silicon): `mlx_vlm` vision probe and `mlx_lm`
text probe both passed for this exact payload.

History: replaces the 2026-04 text-only extraction, which shipped
without the vision tower and failed to load under `mlx_vlm`
("Missing 393 parameters"). Reported in discussion #1 on
majentik/Qwen3.6-35B-A3B-RotorQuant-MLX-4bit.