---

license: apache-2.0

base_model: Jackrong/Qwopus3.6-27B-Coder

tags:

- gguf

- quantization

- asi-evolve

- moq

- qwen3.5

- coder

- research-artifact

- superseded

---

Qwopus3.6-27B-Coder-Predator-Q-ASI-v2

> ⚠️ STATUS (2026-06-24): SUPERSEDED by Qwopus3.6-27B-Coder-GGUF-kaitchup-MoQ-4.0 (MoQ-4.0, 12.4 GiB).

>

> Multi-benchmark validation (HumanEval+ 164, MBPP+ 100, BigCodeBench 50, LCB-30 30) showed no statistically significant difference between v2 and MoQ-4.0 on any benchmark (all McNemar p > 0.31, Bonferroni α=0.0167). The +16% KL-divergence improvement of v2 did not translate to measurable task gains. MoQ-4.0 is preferred because it is 2 GB smaller and requires zero LLM iteration cost to build.

>

> This model is retained as a research artifact and as a baseline for "what LLM-guided recipe search produces." It is not recommended for production use.

What this IS

ASI-Evolved v2 is the result of an LLM-guided recipe search starting from MoQ-4.0 (the kaitchup MoQ recipe applied directly to Qwopus). The ASI-Evolve framework iteratively improved the per-tensor quantization recipe using paired perplexity and KL-divergence against the F16 reference on WikiText-2. After 20 iterations of search:

Quantization recipe (29 tensor-type overrides, 14.4 GiB, ~4.51 BPW):

• L64 (nextn) tensors: bf16 (kaitchup's recipe was missing these — they were defaulting to Q4_K)
• 5 most-sensitive attention layers (L3, L7, L11, L59, L63): IQ4_NL
• Edge-layer ffn_up: IQ4_NL; middle ffn_up: IQ4_XS
• Remainder: kaitchup MoQ-4.0 recipe

Quantitative gains over MoQ-4.0:

• KL-divergence vs F16: 0.034 → 0.0287 (16% improvement)
• PPL ratio: 1.020
• File size: 12.4 → 14.4 GiB (+2 GiB, +16%)

What did NOT improve (Phase 5 multi-benchmark validation):

| Benchmark | MoQ-4.0 | v2 | Δ | McNemar p |

|---|---|---|---|---|

| HumanEval+ base (164) | 92.7% | 93.3% | +0.6pp | 0.65 |

| HumanEval+ plus (164) | 89.6% | 87.8% | -1.8pp | 0.65 |

| MBPP+ base (100) | 92.0% | 93.0% | +1.0pp | 0.56 |

| MBPP+ plus (100) | 79.0% | 78.0% | -1.0pp | 0.56 |

| BigCodeBench (50) | 42.0% | 44.0% | +2.0pp | 0.32 |

| LCB-30 (30) | 50.0% | 53.3% | +3.3pp | 0.72 |

All deltas are within binomial/McNemar noise. The 2 GB size premium and 5+ LLM-call iteration cost were not justified.

!Multi-benchmark comparison

!Summary table

Why v2 exists (research value)

This model demonstrates that:

1. LLM-guided recipe search can find improvements that a human or simple heuristic might miss — specifically, the missing L64 (nextn) tensor override and selective per-layer IQ4_NL application.
2. KL-divergence improvement does not always equal task improvement. v2 is genuinely more F16-like at the logit-distribution level, but this does not translate to better code generation, comprehension, or library use on the 6 tested benchmarks.
3. The MoQ recipe is near-optimal for this model at this BPW range. The kaitchup recipe is a strong local optimum; ASI-Evolve finds a slightly better KL-div point but no measurable downstream gain.

Where v2 MIGHT help (not measured)

• Math reasoning (GSM8K, MATH) — could matter if KL-div wins translate to better arithmetic
• Long-form generation (creative writing, doc generation) — rare-token accuracy might matter
• Multilingual tasks — KL-div gap might be larger for non-English tokens
• Production chat — subjective quality differences not captured by these benchmarks

Future work should test on these dimensions before declaring v2 a clear win in any direction.

Recipe discovery method

1. Start: kaitchup MoQ-4.0 recipe applied directly (Phase 1)
2. Loop: ASI-Evolve (gpt-5.2-codex low reasoning) generates per-tensor quantization recipe variations, scored on paired PPL + KL-div against F16
3. Result after 20 iterations: MoQ-4.0-AttnQSelectiveEdgeNL — 2.9% fitness improvement, 0.5% additional beyond step 1

Reproduction

# Requires the evalplus library
pip install evalplus==0.3.1

# On a 1x L40S 48GB VSI:
python scripts/run_eval.py \
    --benchmark humaneval_plus \
    --model Qwopus3.6-27B-Coder-ASI-MoQ-4.0-14.4GB.gguf \
    --n 164 --max-tokens 1024

# See the full multi-benchmark runner:
# scripts/run_all_evals.sh
# scripts/compare_results.py

Provenance

• Source model: Qwopus3.6-27B-Coder BF16
• Reference: qwopus_f16.gguf (53.8 GB)
• Imatrix: qwopus_imatrix.gguf (14 MB)
• Base logits: base_logits_0.dat (5.9 GB, 50 chunks × 512 ctx)
• Recipe: recipe.txt (29 tensor-type overrides)
• Built: 2026-06-23, VSI predator-vsi-05, 1x L40S 48GB
• Validation date: 2026-06-24 (Phase 5 multi-benchmark test)
• SHA256: 63ab13aab9c1078a782ad02f725ee69f7f3c5d398b30759c814afef5d56272fa

Honest assessment

ASI-Evolve produced a model that is more F16-like at the logit distribution level but not measurably better at coding tasks. The 2 GB cost and 5+ LLM-call iteration cost were not justified. The Pareto-optimal choice is MoQ-4.0 (smaller, equivalent).

The reason this is still worth shipping as a research artifact: it is the closest we can get to BF16 quality in 14.4 GB, and may matter for tasks not captured by these benchmarks. Future work could test on additional benchmarks.

License

Apache 2.0 (inherited from base model)