---
library_name: mlx
license: apache-2.0
pipeline_tag: image-text-to-text
tags:
- mlx
- apple-silicon
- speculative-decoding
- qwen
- qwen3
- qwen3-next
- mtp
- mtplx
- local-ai
- vision-language
- multimodal
base_model:
- Youssofal/Qwen3.6-27B-MTPLX-Optimized-Speed
- Qwen/Qwen3.6-27B
---

# Qwen3.6-27B MTPLX Optimized Speed (Vision-Enabled)

A unified local model that both codes at the Qwen3.6-27B level and processes images/videos, packaged for Apple Silicon via **MTPLX** native speculative decoding with Multi-Token-Prediction (MTP).

This repository restores the multimodal vision tower that is typically stripped in text-only MLX/MTPLX conversions.

## Architecture

- **Body**: 4-bit affine, group size 64 (from [Youssofal/Qwen3.6-27B-MTPLX-Optimized-Speed](https://huggingface.co/Youssofal/Qwen3.6-27B-MTPLX-Optimized-Speed))
- **MTP sidecar (mtp/weights.safetensors)**: Prequantized MTP heads with a 3-bit affine group-64 draft-only LM head.
- **Vision tower (vision_tower.safetensors, 333 tensors, bf16 unquantized)**: Extracted verbatim from the official checkpoint's own `model.visual.*` weights (shards 7 and 8 of [Qwen/Qwen3.6-27B](https://huggingface.co/Qwen/Qwen3.6-27B)) and renamed to the `vision_tower.*` prefix for MTPLX native VLM compatibility (~879 MB).

## Usage (MTPLX)

```bash
mtplx start --model <path-to-this-model-dir> --port 8092 \
  --chat-template-path <path-to-this-model-dir>/chat_template.jinja
```

Serves an OpenAI-compatible endpoint supporting both text and image input payloads via `POST /v1/chat/completions`.

The MTPLX CLI defaults to `mtp_history_policy=committed`, which this model
requires for healthy deep-position acceptance. The model's recommended
draft sampler is temperature 0.7 (see `recommended_draft_sampler` in
`mtplx_runtime.json`); pass `--draft-temperature 0.7` to match.

## Performance (measured)

Apple M5 Pro 64 GB, MTPLX 1.0.4, temperature 0.6, `mtp_history_policy=committed`,
`--draft-temperature 0.7` (model recommended), thinking OFF, warm (4 warmup
prompts discarded), 8 measured prompts from the calibration_coding suite,
max_tokens=192, median tok/s.

| Depth | tok/s (wall-clock e2e) | tok/s (decode-only) | speedup vs AR (e2e) | acceptance pos1/2/3 |
|---|---|---|---|---|
| AR (`--no-mtp`) | 13.7 | 14.5 | 1.00x | - |
| MTP depth 1 | 19.7 | 20.9 | 1.44x | 0.923 |
| MTP depth 2 | 19.9 | 20.8 | 1.46x | 0.950 / 0.753 |
| MTP depth 3 | 18.1 | 19.7 | 1.32x | 0.900 / 0.759 / 0.648 |

Two throughput definitions are reported:

- **tok/s (wall-clock e2e)** = `generated_tokens / total_elapsed`, includes
  prompt prefill. The real end-to-end speed and the honest denominator for
  the speedup ratio.
- **tok/s (decode-only)** = `generated_tokens / decode_elapsed`, excludes
  prefill.

Acceptance is stable across runs: greedy draft (temp 0.0) gives
0.874/0.740/0.636, recommended draft (temp 0.7) gives 0.900/0.759/0.648,
both within noise. Vision graft has zero acceptance cost relative to the
text-only source (Youssofal/Qwen3.6-27B-MTPLX-Optimized-Speed measures
identical acceptance 0.874/0.740/0.636 under the same config).

Earlier baked `mtplx_runtime.json` figures (acceptance 1.0/0.98/0.94, ~63
tok/s) were measured on Apple M5 Max 128 GB and are not reproducible on M5
Pro 64 GB due to lower memory bandwidth; they are preserved in the
`historical_m5max` field of `mtplx_runtime.json` for traceability. The
vision tower is lazy-loaded (only when a request carries an image), so
text-only throughput is unaffected by it. Numbers are indicative, not a
benchmark suite.

## Sources And Attribution

| Component | Source | Revision / License |
|---|---|---|
| Base model | `Qwen/Qwen3.6-27B` | Apache-2.0 |
| Quantized Text/MTP repo | `Youssofal/Qwen3.6-27B-MTPLX-Optimized-Speed` | Apache-2.0 |
| MTPLX conversion | `youssofal/mtplx` | Apache-2.0 |