---
language: en
pipeline_tag: text-generation
library_name: mlx
license: apache-2.0
base_model: Jackrong/Qwopus3.6-27B-Coder
tags:
- mlx
- mtplx
- mtp
- speculative-decoding
- apple-silicon
- qwopus
- code
- agent
---

# Qwopus3.6-27B-Coder · MTPLX 4-bit Speed

The first [MTPLX](https://github.com/youssofal/MTPLX) build of **Qwopus3.6-27B-Coder** — Jackrong's agentic-coding fine-tune of Qwopus3.6-27B-v2 (repo-level coding, multi-turn tool orchestration, 67.0% SWE-bench Verified in no-thinking mode), with native multi-token-prediction speculative decoding on Apple Silicon. No external drafter, exact rejection sampling. 8-bit sibling: [nom666/Qwopus3.6-27B-Coder-MTPLX-8bit-Quality](https://huggingface.co/nom666/Qwopus3.6-27B-Coder-MTPLX-8bit-Quality).

Forged with `mtplx forge build` from the original BF16 [Jackrong/Qwopus3.6-27B-Coder](https://huggingface.co/Jackrong/Qwopus3.6-27B-Coder):

- **Body:** flat 4-bit MLX affine quantization, group size 64
- **MTP head:** preserved in BF16 (`mtp_policy: keep_bf16`), packed as `mtp.safetensors` sidecar
- **Size:** ~15 GB
- **Calibrated MTP contract** included (`mtplx_runtime.json`)

## Measured performance (Apple M5 Max, 128 GB, MTPLX 1.0.3)

| Mode | Decode | Acceptance by depth |
|---|---|---|
| **MTP depth 2 (winner)** | **57.3 tok/s** | 97% / 90% |

The Coder's MTP head was trained with draft=2, which shows: depth-2 acceptance is exceptional and depth 2 beats depth 3. Verification suite: `long-code-uncapped`, 2048-token budget. Use `--depth 2` when serving.

## Usage

```bash
brew install youssofal/mtplx/mtplx   # or pipx install mtplx
mtplx pull nom666/Qwopus3.6-27B-Coder-MTPLX-4bit-Speed
mtplx quickstart --model nom666/Qwopus3.6-27B-Coder-MTPLX-4bit-Speed \
  --depth 2 --paged-kv-quantization q8 --batching-preset agent --reasoning off
```

Serves OpenAI-compatible (`/v1/chat/completions`) and Anthropic-compatible (`/v1/messages`) endpoints with warm-prefix KV reuse, SSD session cache, continuous batching, and vision support. Full 262144-token context; only 16 of 64 layers carry KV (hybrid Gated DeltaNet architecture).

## Notes

- This fine-tune targets no-thinking agentic use — pair with `--reasoning off`.
- Runtime contract tier is `forge-local`: verified on the forging machine (M5 Max).
- Quantized with MTPLX 1.0.3 Forge. All credit for the fine-tune to Jackrong; base Qwen3.6-27B (Apache-2.0).