---
license: other
license_name: qwen
license_link: https://huggingface.co/Qwen/Qwen3.6-27B/blob/main/LICENSE
base_model:
- KyleHessling1/Qwopus3.6-27B-Fusion-BF16
base_model_relation: quantized
library_name: mlx
pipeline_tag: image-text-to-text
tags:
- mlx
- mxfp8
- qwen3.5
- experimental
- mlx-vlm
- vision
- mtplx
- speculative-decoding
- multi-token-prediction
---

# Qwopus3.6-27B-Fusion-MXFP8-Vision-MTPLX

Local MLX conversion of [KyleHessling1/Qwopus3.6-27B-Fusion-BF16](https://huggingface.co/KyleHessling1/Qwopus3.6-27B-Fusion-BF16), pinned at `c4196377f197b510b4821f246c81a3595c781ecf`. The Fusion weights were reconstructed from the two pinned BF16 parent repositories with the publisher's exact layer-weighted formula; all 1,199 source tensors passed an independent exact check before quantization.

## Format

- Language body: MXFP8, 8-bit, group size 32
- Vision tower: 333 source BF16 tensors; two direct `mlx_vlm` image smokes passed
- Architecture: dense Qwen3.5, 64 language layers, 262,144-token configured context
- Tokenizer and chat template: preserved from the pinned source lineage

## MTPLX

This self-contained package adds the source model's native 15-tensor BF16 MTP sidecar. The sidecar was verified tensor-exact against the pinned Fusion reconstruction. Depth 3 is recommended; local 192-token tuning measured 1.47x (23.46 vs 15.92 tok/s) versus autoregressive decoding.

```bash
mtplx inspect --require-mtp <local-path-or-repo-id>
mtplx serve --model <local-path-or-repo-id> --generation-mode mtp --mtp-depth 3 --reasoning-mode off
```

## Local qualification

- Structural/reproduction release qualifier: 60/60 checks passed
- Strict text load: passed; executable code smoke: 3/3
- Native and parallel tool-call gates: 12/12
- Long-context needle: passed
- Hard behavioral suite: **4/10**

The hard-suite score is an important quality caveat. This is a research-preview merge and runtime-qualified conversion, not a claim that every coding or reasoning edge case passes. The source publisher likewise labels Fusion experimental and not fully evaluated. The configured 262K context was not exhaustively qualified here.

## Basic use

```bash
python -m mlx_vlm.generate --model <local-path-or-repo-id> --image image.jpg --prompt "Describe this image." --max-tokens 256
```

Review the [Fusion source card](https://huggingface.co/KyleHessling1/Qwopus3.6-27B-Fusion-BF16) and the Qwen license before use. This conversion does not change the source model's safety characteristics.