---
library_name: transformers
base_model:
- Jackrong/Qwopus3.6-27B-v2
license: apache-2.0
license_name: apache-2.0
license_link: https://www.apache.org/licenses/LICENSE-2.0
language:
- en
- zh
- es
- ru
- ja
pipeline_tag: image-text-to-text
tags:
- gguf
- llama.cpp
- auto-round
- q4_k_m
- image-text-to-text
- vision
- multimodal
- qwen3_6
- quantized
- text-generation-inference
- transformers
- unsloth
- conversational
- reasoning
- chain-of-thought
- tool-use
- function-calling
---

# Qwopus3.6-27B-v2 - AutoRound Q4_K_M GGUF

> Quantized with [Intel AutoRound](https://github.com/intel/auto-round) (SignRound optimization) and exported to GGUF for [llama.cpp](https://github.com/ggml-org/llama.cpp).

## Files

| File | Method | Description |
|------|--------|-------------|
| `Qwopus3.6-27B-v2-Q4_K_M-iter0.gguf` | AutoRound RTN | Round-to-nearest, no optimization |
| `Qwopus3.6-27B-v2-Q4_K_M-iter100.gguf` | AutoRound iter=100 | 100 rounds of SignRound SGD optimization |

## Generation Commands

### Iter 0 (RTN, fast)
```bash
AR_DISABLE_COPY_MTP_WEIGHTS=1 auto-round \
    --model Jackrong/Qwopus3.6-27B-v2 \
    --scheme W4A16 \
    --format "gguf:q4_k_m" \
    --iters 0 \
    --output_dir ./output
```

### Iter 100 (SignRound optimized, ~1h on RTX PRO 6000)
```bash
AR_DISABLE_COPY_MTP_WEIGHTS=1 auto-round \
    --model Jackrong/Qwopus3.6-27B-v2 \
    --scheme W4A16 \
    --format "gguf:q4_k_m" \
    --iters 100 \
    --output_dir ./output
```

## KL Divergence Evaluation

Tested on wikitext-2-raw-v1 (145 chunks, ctx=2048, batch=2048, ngl=99, Blackwell CUDA).

| Metric | Q8_0 (Ref) | Jackrong Q4_K_M | AR iter=0 | AR iter=100 |
|--------|------------|-----------------|-----------|-------------|
| PPL | 6.180 | 6.380 | 6.344 | **6.207** |
| PPL vs Q8_0 | — | +3.24% | +2.66% | **+0.44%** |
| Mean KL Divergence | — | 0.0494 | 0.0483 | **0.0435** |
| Median KL Divergence | — | 0.0083 | 0.0080 | **0.0075** |
| RMS Δp | — | 5.51% | 5.45% | **5.13%** |
| Same Top-p | — | 93.57% | 93.61% | **93.93%** |
| Model Size | 27 GB | 16 GB | 16 GB | 16 GB |

**Key finding:** AutoRound iter=100 achieves near-lossless Q4_K_M quantization with only 0.44% PPL degradation vs Q8_0, significantly outperforming standard llama.cpp Q4_K_M (3.24% degradation).

### KL Divergence Commands

```bash
# Step 1: Generate base logits (Q8_0)
llama-perplexity -m Q8_0.gguf -f wiki.test.raw \
    --save-all-logits q8_0_base.logits -ngl 99 -c 2048 -b 2048

# Step 2: Compute KL divergence
llama-perplexity -m Q4_K_M.gguf -f wiki.test.raw \
    --kl-divergence --kl-divergence-base q8_0_base.logits \
    -ngl 99 -c 2048 -b 2048
```

---

# Original README from Jackrong/Qwopus3.6-27B-v2-GGUF

(Content below is from the original model card.)

---

<div style="font-family: -apple-system, BlinkMacSystemFont, 'Segoe UI', Roboto, sans-serif; border: 1px solid #cbd5e1; border-radius: 16px; box-shadow: 0 10px 15px -3px rgba(0, 0, 0, 0.05), 0 4px 6px -2px rgba(0, 0, 0, 0.05); overflow: hidden; background: #ffffff; margin-bottom: 30px;">
  <div style="background: linear-gradient(135deg, #7c3aed 0%, #4c1d95 100%); padding: 24px; color: white;">
    <div style="display: flex; align-items: center; justify-content: space-between; flex-wrap: wrap; gap: 10px;">
      <h1 style="margin: 0; font-size: 26px; font-weight: 800; display: flex; align-items: center; gap: 12px; color: white; border: none;">🪐 Qwopus3.6-27B-v2</h1>
      <span style="background: #10b981; color: white; font-size: 11px; font-weight: 700; padding: 4px 10px; border-radius: 20px; text-transform: uppercase; letter-spacing: 0.5px;">SFT Release</span>
    </div>
    <p style="margin: 8px 0 0 0; font-size: 14px; color: #ddd6fe; font-weight: 500;">Reasoning-Enhanced Dense Language Model Fine-Tuned on Qwen3.6-27B</p>
  </div>
  <div style="display: flex; gap: 8px; flex-wrap: wrap; padding: 12px 24px; background: #f8fafc; border-bottom: 1px solid #e2e8f0;">
    <span style="background: #f3e8ff; color: #6b21a8; font-size: 11px; font-weight: 700; padding: 4px 10px; border-radius: 20px; border: 1px solid #e9d5ff;">🧬 Trace Inversion & Negentropy</span>
    <span style="background: #dbeafe; color: #1e40af; font-size: 11px; font-weight: 700; padding: 4px 10px; border-radius: 20px; border: 1px solid #bfdbfe;">🧠 27B Parameters</span>
    <span style="background: #e0f2fe; color: #0369a1; font-size: 11px; font-weight: 700; padding: 4px 10px; border-radius: 20px; border: 1px solid #bae6fd;">🔥 3-Stage Curriculum SFT</span>
    <span style="background: #d1fae5; color: #065f46; font-size: 11px; font-weight: 700; padding: 4px 10px; border-radius: 20px; border: 1px solid #a7f3d0;">🛠️ Vision & Tool-use Support</span>
  </div>
  <div style="padding: 24px; display: flex; flex-direction: column; gap: 20px;">
    <div style="background: #f5f3ff; border-left: 5px solid #7c3aed; padding: 16px; border-radius: 0 8px 8px 0;">
      <h3 style="margin: 0 0 8px 0; font-size: 15px; color: #6d28d9; font-weight: 700; display: flex; align-items: center; gap: 6px;"><span>💡</span> What is Qwopus3.6-27B-v2?</h3>
      <p style="margin: 0; font-size: 13px; color: #334155; line-height: 1.6;">🪐 <b>Qwopus3.6-27B-v2</b> is a reasoning-enhanced dense language model built on <b>Qwen3.6-27B-base</b> via the <b>Trace Inversion & Negentropy framework</b>. It's trained on ~14K highly informative samples derived from Claude Opus 4.x reasoning traces, following a 3-stage curriculum spanning context expansion (32K→128K), reasoning strengthening, and tool-use integration.</p>
    </div>
  </div>
</div>