majentik

/

Qwen3.6-35B-A3B-RotorQuant-MLX-4bit

like 5

Thanks for the report, and sorry for the trouble — you found a real packaging bug, now fixed.

What was wrong: this repo (and its 11 MLX siblings) originally shipped only the quantized language tower of Qwen3.6-35B-A3B. The vision tower was missing entirely, so mlx_vlm correctly refused to load it ("Missing 393 parameters"). The card's multimodal claim and mlx_vlm snippet were wrong.

What changed (2026-07-04): all 12 Qwen3.6-35B-A3B-{RotorQuant,TurboQuant}-MLX-{2..8}bit repos were re-published as full-tower packs, converted with mlx-vlm 0.6.3 from upstream rev 995ad96:

• language tower quantized (per-repo bit width, affine, group size 64)
• vision tower included in BF16 (unquantized, 333 tensors)

Every variant was smoke-tested on this exact payload before publishing — mlx_vlm image inference and mlx_lm text-only both pass (details in each repo's PROVENANCE.md). Load with mlx-vlm >= 0.6:

from mlx_vlm import load, generate
from mlx_vlm.prompt_utils import apply_chat_template

model, processor = load("majentik/Qwen3.6-35B-A3B-RotorQuant-MLX-4bit")
prompt = apply_chat_template(processor, model.config, "Describe this image.", num_images=1)
output = generate(model, processor, prompt, image=["path/to/image.jpg"], max_tokens=256)
print(output.text)

Note the new payloads are slightly larger than before (this repo: ~19 GB) because the BF16 vision tower is now included. Video input is architecturally supported upstream but wasn't part of our smoke gate, so treat it as untested. If anything else looks off, please reopen.

Closing — fixed by the 2026-07-04 full-tower republish (see above). Reopen if anything still fails.