---
license: apache-2.0
base_model: Jackrong/Qwopus3.6-27B-Coder
tags:
  - gguf
  - quantization
  - moq
  - mixture-of-quantizations
  - qwen3.5
  - coder
  - code
  - verified
---

# Qwopus3.6-27B-Coder-GGUF-Predator-Q (MoQ-4.0)

**Mixture-of-Quantizations (MoQ) 4.0 BPW GGUF for Qwopus3.6-27B-Coder.**

**Canonical Predator-Q winner as of 2026-06-24** — verified on 6 benchmarks (HumanEval+ 164, MBPP+ 100, BigCodeBench 50, LCB-30 30). See "Multi-benchmark validation" section below for the full comparison vs. ASI-Evolved v2.

## Summary

| Metric | Value |
|---|---|
| File size | **12.6 GB** |
| Effective BPW | 4.04 |
| WikiText-2 PPL (Q) | 6.18 (vs F16 base 6.02) |
| Mean PPL(Q)/PPL(base) | 1.026 |
| Mean KL divergence (vs F16) | 0.034 |
| Same top p (vs F16) | 91.6% |
| RMS Δp | 5.43% |
| **HumanEval+ base pass@1** | **92.7% (152/164)** |
| **HumanEval+ plus pass@1** | **89.6% (147/164)** |
| **MBPP+ base pass@1** | **92.0% (92/100)** |
| **MBPP+ plus pass@1** | **79.0% (79/100)** |
| **BigCodeBench pass@1** | **42.0% (21/50)** |
| **LCB-30 pass@1** | **50.0% (15/30)** |

## Multi-benchmark validation (2026-06-24, Phase 5)

**This is the canonical winner.** A more complex alternative (ASI-Evolved v2) was built and tested on the same 4 multi-task benchmarks. Result: **no statistically significant difference** on any benchmark (all McNemar p > 0.31, Bonferroni-corrected α = 0.0167). MoQ-4.0 is preferred because it is **2 GB smaller and requires zero LLM iteration cost** to build.

![Multi-benchmark comparison](benchmark_comparison_bars.png)

| Benchmark | MoQ-4.0 (12.4 GiB) | ASI-Evolved v2 (14.4 GiB) | Δ | McNemar p |
|---|---|---|---|---|
| HumanEval+ base (164) | 152 (92.7%) | 153 (93.3%) | +0.6pp | 0.65 |
| HumanEval+ plus (164) | 147 (89.6%) | 144 (87.8%) | -1.8pp | 0.65 |
| MBPP+ base (100) | 92 (92.0%) | 93 (93.0%) | +1.0pp | 0.56 |
| MBPP+ plus (100) | 79 (79.0%) | 78 (78.0%) | -1.0pp | 0.56 |
| BigCodeBench (50) | 21 (42.0%) | 22 (44.0%) | +2.0pp | 0.32 |
| LCB-30 (30) | 15 (50.0%) | 16 (53.3%) | +3.3pp | 0.72 |

![Summary table](summary_table.png)

The v2 model has a 16% lower KL-divergence against F16 (0.0287 vs 0.034) but **does not produce measurably better outputs on any of the 6 code tasks tested**. The 2 GB file size savings of MoQ-4.0 are pure upside.

## Pareto frontier

![Pareto frontier](pareto_size_vs_kldiv.png)
![Pareto size vs capability](pareto_size_vs_capability.png)

MoQ-4.0 sits at the **Pareto-optimal corner**: smaller than all baseline quantization recipes, with capability equivalent to (or statistically tied with) the larger alternatives.

## Comparison to Baselines

| Build | File Size | BPW | KL-div | Same top p | PPL ratio | LCB-30 |
|-------|-----------|-----|--------|------------|-----------|--------|
| Q4_K_M | 15.4 GB | 4.69 | 0.019 | 93.6% | 1.009 | 50.0% |
| 4A BEST | 16.98 GB | 5.05 | 0.018 | 93.7% | 1.012 | 50.0% |
| ASI-Evolved v2 | 14.4 GiB | 4.51 | 0.0287 | 92.55% | 1.020 | 53.3% |
| **This (MoQ-4.0)** | **12.6 GB** | **4.04** | **0.034** | **91.6%** | **1.026** | **50.0%** |

**Net result:** Same HumanEval+/MBPP+/LCB-30 capability as Q4_K_M and 4A BEST, with **18% smaller file size** than Q4_K_M and **26% smaller** than 4A BEST. ASI-Evolved v2 has a marginal KL-div win that did not translate to task performance.

## What is MoQ?

MoQ (Mixture of Quantizations) is a search-based mixed-precision quantization method by Waleed Ahmad. It estimates per-tensor "elasticity" — sensitivity to low precision — then assigns each tensor a different Q-type to minimize quality loss per bit spent. This produces non-uniform recipes like:

- `ffn_down` → IQ3_S (aggressive — most redundant tensor)
- `ffn_gate` → IQ3_S (aggressive)
- `ffn_up` → IQ4_XS (slightly more conservative)
- `attn_qkv` (most layers) → IQ4_XS
- `attn_qkv` (special every-4th full-attention layers) → split: `attn_q`=IQ4_NL, `attn_k`=bf16, `attn_v`=bf16, `attn_output`=Q5_K
- `token_embd` → Q4_K
- Norms → bf16

The recipe was applied with **importance matrix (imatrix)** calibration, which makes IQ-quant decisions properly importance-aware. Without imatrix, the same recipe produces worse KL-div (0.059 vs 0.034).

## Recipe Source

The per-tensor recipe used here is `MoQ_qwen3.6-27b_tensors_4.0.txt` from [kaitchup/Qwen3.6-27B-GGUF-MoQ](https://huggingface.co/kaitchup/Qwen3.6-27B-GGUF-MoQ). Since Qwopus3.6-27B-Coder uses the same Qwen3.5 architecture (qwen35, 64 layers, 851 tensors), the recipe transfers directly without modification.

## Build Details

```bash
# 1. Generate imatrix (WikiText-2 test set, 100 chunks)
llama-imatrix -m qwopus_f16.gguf -f wikitext-2-raw/wiki.test.raw \
  -o qwopus_imatrix.gguf -ngl 50 --chunks 100

# 2. Quantize with MoQ recipe + imatrix
llama-quantize --imatrix qwopus_imatrix.gguf \
  --tensor-type-file MoQ_qwen3.6-27b_tensors_4.0.txt \
  qwopus_f16.gguf qwopus_moq_4.0.gguf Q4_K
```

## Evaluation Details

- **WikiText-2 KL-div:** `llama-perplexity -c 512 -b 512 --kl-divergence --kl-divergence-base base_logits_0.dat --chunks 50 --seed 1337`
- **HumanEval+ / MBPP+:** evalplus library, untrusted_check with base+plus variants, max_tokens=1024, temperature=0
- **BigCodeBench:** custom harness using unittest test cases, 30s timeout per problem
- **LCB-30:** LiveCodeBench easy subset, 30 problems × 3 attempts each (pass@1)
- **Statistical test:** McNemar's test (paired, binary outcomes), Bonferroni correction across 6 benchmarks
- **Base logits:** F16 GGUF reference logits saved once and reused for all variant evaluations

## Known Limitations

- **Code-specific tasks:** Tested on 6 coding benchmarks. Performance on math/reasoning/long-form generation not measured.
- **Long context:** Not stress-tested beyond LCB-30's 4096-token context. Qwopus's full 262K context window requires the 4D-RoPE / Mamba-SSM hybrid architecture to work correctly — preserved here since all norms/embeddings are at full precision.
- **KL-div is higher than Q4_K_M** (0.034 vs 0.019). This is the cost of smaller file size with mixed-precision quantization. Same top token agreement remains high (91.6%).

## Provenance

- **Base model:** [Jackrong/Qwopus3.6-27B-Coder](https://huggingface.co/Jackrong/Qwopus3.6-27B-Coder) (BF16)
- **Recipe source:** [kaitchup/Qwen3.6-27B-GGUF-MoQ](https://huggingface.co/kaitchup/Qwen3.6-27B-GGUF-MoQ) `MoQ_qwen3.6-27b_tensors_4.0.txt`
- **Methodology reference:** Waleed Ahmad's MoQ ([w-ahmad](https://huggingface.co/w-ahmad)); kaitchup substack [article](https://kaitchup.substack.com/p/moq-ggufs-and-gsq-low-bit-ggufs-are)
- **Calibration data:** WikiText-2 test set (~80k lines), chunks=100
- **Multi-benchmark validation date:** 2026-06-24 (Phase 5)
- **Build date:** 2026-06-23
- **SHA256:** `587840e75895199e5ad771bfa7dfd9682f6d85ae295ad00001b78adb485c52c1`

## Reproduce

```bash
# Clone recipes from kaitchup
wget https://huggingface.co/kaitchup/Qwen3.6-27B-GGUF-MoQ/resolve/main/MoQ_qwen3.6-27b_tensors_4.0.txt

# Download Qwopus BF16 shards and convert to F16 GGUF
# (see Jackrong/Qwopus3.6-27B-Coder for conversion instructions)

# Generate imatrix
llama-imatrix -m qwopus_f16.gguf -f wikitext-2-raw/wiki.test.raw \
  -o qwopus_imatrix.gguf -ngl 50 --chunks 100

# Quantize
llama-quantize --imatrix qwopus_imatrix.gguf \
  --tensor-type-file MoQ_qwen3.6-27b_tensors_4.0.txt \
  qwopus_f16.gguf qwopus_moq_4.0.gguf Q4_K

# Validate
./llama-perplexity -m qwopus_moq_4.0.gguf -f wikitext-2-raw/wiki.test.raw \
  -c 2048 --chunks 30
```

## License

Apache 2.0 (inherited from base model)