---

license: apache-2.0

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

---

⚠️ THIS REPO HAS BEEN RENAMED

This repository is deprecated. The contents have been moved to:

Fredred89/Qwopus3.6-27B-Coder-GGUF-ASI-MoQ-4.0

Why the rename: The "Predator-Q" branding implied novel work. While ASI-Evolved v2 does include 2 real improvements over kaitchup's recipe (L64 nextn bf16, selective IQ4_NL on 5 attention layers), the new repo name properly attributes the source as kaitchup's MoQ recipe + ASI-Evolve iteration.

What we actually did:

1. Started with kaitchup's MoQ-4.0 recipe
2. Ran 20 iterations of ASI-Evolve (gpt-5.2-codex) to find recipe improvements
3. Found L64 (nextn) tensors should be bf16, selective IQ4_NL on 5 most-sensitive attention layers
4. Validated with multi-benchmark testing (HumanEval+ 164, MBPP+ 100, BigCodeBench 50, LCB-30 30)

Result: v2 achieves 16% better KL-divergence (0.034 → 0.0287) but 0% task improvement over MoQ-4.0 (all McNemar p > 0.31, Bonferroni α = 0.0167). The 2 GB file size premium and 5+ LLM-call iteration cost were not justified.

Recommendation: Use Fredred89/Qwopus3.6-27B-Coder-GGUF-kaitchup-MoQ-4.0 instead — same task performance, 2 GB smaller, simpler recipe.

---

Attribution

• Starting point: kaitchup's MoQ recipe at kaitchup/Qwen3.6-27B-GGUF-MoQ
• ASI-Evolve framework: GAIR-NLP/ASI-Evolve
• LLM used for recipe search: gpt-5.2-codex (low reasoning)
• Base model: Jackrong/Qwopus3.6-27B-Coder

The new repo (Fredred89/Qwopus3.6-27B-Coder-GGUF-ASI-MoQ-4.0) contains the same GGUF plus full multi-benchmark validation results showing the 0% task gain from ASI iteration.