---
license: apache-2.0
library_name: mlx
tags:
- mtplx
- mlx
- qwen3_5
- qwen3.6
- image-text-to-text
- speculative-decoding
- multi-token-prediction
- Qwen/Qwen3.6-27B
---

# Qwen3.6 27B MTPLX Optimized Quality

This is the MTPLX **Optimized Quality** artifact for Qwen3.6 27B.

Policy:

- Target trunk: true flat 8-bit MLX affine quantization, group64.
- MTP auxiliary tensors (`mtp.fc`, norms, scales/biases, and non-linear tensors) are preserved in BF16 where applicable.

This is intended as the higher-quality sibling to `Qwen3.6-27B-MTPLX-Optimized-Speed`. It favors the Flat8 target and calibrated INT8 proposal sidecar instead of the smaller speed-focused artifact.

## MTPLX

```bash
mtplx start --model Youssofal/Qwen3.6-27B-MTPLX-Optimized-Quality
```

The artifact includes `mtplx_runtime.json` and `mtp/weights.safetensors`, so MTPLX can inspect and route it through the native Qwen MTP backend while generic MLX vision loaders only glob the base model shards.

## Local bakeoff snapshot

Measured on the local M5 Max max-fan Flappy 2k depth-3 bakeoff:

| Metric | Value |
|---|---:|
| Decode TPS | 33.63 |
| Acceptance D1/D2/D3 | 95.6% / 85.3% / 74.1% |
| Verify ms/call | 88.1 ms |
| Peak memory | 27.62 GiB |

This row is a local release-readiness check, not a broad public hardware claim.

## Provenance

- Base model: `Qwen/Qwen3.6-27B`
- MTPLX staging manifest: `mtplx_upload_manifest.json`

## Vision compatibility

This revision keeps the optimized MTPLX draft sidecar at `mtp/weights.safetensors` so generic VLM loaders see only the normal Qwen vision/text weight shards at the repository root. The base trunk is the MLX 8-bit Qwen3.6 vision layout; MTPLX reads the draft sidecar through `config.json`.