---
license: apache-2.0
base_model: Jackrong/Qwopus3.6-35B-A3B-Coder
tags:
  - gguf
  - quantized
  - apex
  - apex-mtp
  - moe
  - mixture-of-experts
  - qwen3
  - speculative-decoding
  - self-speculative
  - mtp
  - vlm
  - vision
  - coder
---

<!-- apex-banner-v2 -->
<div style="background-color: #f59e0b; color: white; padding: 20px; border-radius: 10px; text-align: center; margin: 20px 0;">
<h2 style="color: white; margin: 0 0 10px 0;">⚡ Each donation = another big MoE quantized</h2>
<p style="font-size: 18px; margin: 0 0 15px 0;">I host <b>30+ free APEX MoE quantizations</b> as independent research. My only local hardware is an <b>NVIDIA DGX Spark</b> (122 GB unified memory), enough for ~30-50B-class MoEs, but <b>bigger ones (200B+) require rented compute</b> on H100/H200/Blackwell, typically $20-100 per quant.<br>If APEX quants are useful to you, your support directly funds those bigger runs.</p>
<p style="font-size: 20px; margin: 0;">
<a href="https://www.patreon.com/cw/mudler" style="color: white; text-decoration: underline;">🎉 Patreon (Monthly)</a> &nbsp;|&nbsp;
<a href="https://www.buymeacoffee.com/mudler" style="color: white; text-decoration: underline;">☕ Buy Me a Coffee</a> &nbsp;|&nbsp;
<a href="https://github.com/sponsors/mudler" style="color: white; text-decoration: underline;">⭐ GitHub Sponsors</a>
</p>
</div>

# Qwopus3.6-35B-A3B-Coder — APEX-MTP GGUF

**APEX (Adaptive Precision for EXpert Models)** quantizations of [Jackrong/Qwopus3.6-35B-A3B-Coder](https://huggingface.co/Jackrong/Qwopus3.6-35B-A3B-Coder), with the model's **MTP (multi-token prediction) head bundled** for in-the-box self-speculative decoding.

**Brought to you by the [LocalAI](https://github.com/mudler/LocalAI) team** | [APEX Project](https://github.com/mudler/apex-quant) | [Technical Report](https://github.com/mudler/apex-quant/blob/main/paper/APEX_Technical_Report.pdf)

## What's different from the plain APEX repo?

This model ships a real **MTP head**, and these GGUFs bundle it alongside the trunk in a single file (via [llama.cpp PR #22673](https://github.com/ggml-org/llama.cpp/pull/22673)). With a recent llama.cpp you can enable self-speculative decoding from just this one file — no separate draft model:

```bash
llama-server -m Qwopus3.6-35B-A3B-Coder-APEX-MTP-I-Balanced.gguf --draft-mtp
```

The non-MTP version is at [mudler/Qwopus3.6-35B-A3B-Coder-APEX-GGUF](https://huggingface.co/mudler/Qwopus3.6-35B-A3B-Coder-APEX-GGUF) — slightly smaller, no self-spec.

## MTP draft head precision

The bundled MTP head (`blk.40.*` including `nextn.*`) is quantized to **Q8_0** (near-lossless) on every tier, keeping draft accuracy high for a good spec-decode acceptance rate at a modest size cost. The MTP head is not imatrix-calibrated (imatrix forward passes only activate the trunk), so it uses static Q8_0.

## Available Files

| File | Profile | Best For |
|------|---------|----------|
| Qwopus3.6-35B-A3B-Coder-APEX-MTP-I-Balanced.gguf | I-Balanced | Best overall + self-spec |
| Qwopus3.6-35B-A3B-Coder-APEX-MTP-I-Quality.gguf | I-Quality | Highest quality with imatrix + self-spec |
| Qwopus3.6-35B-A3B-Coder-APEX-MTP-Quality.gguf | Quality | Highest quality (no imatrix) |
| Qwopus3.6-35B-A3B-Coder-APEX-MTP-Balanced.gguf | Balanced | General purpose |
| Qwopus3.6-35B-A3B-Coder-APEX-MTP-I-Compact.gguf | I-Compact | Consumer GPUs + self-spec |
| Qwopus3.6-35B-A3B-Coder-APEX-MTP-Compact.gguf | Compact | Consumer GPUs |
| Qwopus3.6-35B-A3B-Coder-APEX-MTP-I-Mini.gguf | I-Mini | Smallest viable + self-spec |
| mmproj.gguf | Vision projector | Required for image understanding |

## Architecture

- **Base**: Qwopus3.6-35B-A3B-Coder (Qwen3_5MoeForConditionalGeneration, Qwen3.6-35B-A3B)
- **Layers**: 40 trunk + 1 MTP (bundled) · **Experts**: 256 routed + 1 shared (8 active)
- **Vision**: Built-in vision encoder (mmproj included)
- **Calibration**: v1.3 diverse dataset

## Credits

APEX by the [LocalAI](https://github.com/mudler/LocalAI) team. MTP support: llama.cpp PR #22673. Built on [llama.cpp](https://github.com/ggerganov/llama.cpp). Base model by [Jackrong](https://huggingface.co/Jackrong).