Acnoryx AI Research Bundle
Overview
- Base model: Qwen/Qwen3.5-0.8B
- Model size: 0.8b
- Research quantizations: Q3_K_M, IQ3_M, Q2_K, IQ2_M, IQ2_XS, IQ2_XXS, IQ1_M, IQ1_S
- Purpose: evaluate quality vs. size trade-offs below the production threshold
Notes
- IQ1/IQ2 formats require an importance matrix (imatrix).
- These files are more experimental than the release bundle.
- Production-facing use should prefer the release bundle.
- If prompting in Vietnamese, write with full accents for best consistency.
Evaluation Snapshot
Research GGUFs were continued from the existing results and merged with the latest rerun on the same curated 58-question bilingual benchmark.
| Quant | Think | No-Think | Avg | Status |
|---|
| Q3_K_M | 74.1% | 72.4% | 73.2% | Best current research quant |
| IQ3_M | 60.3% | 60.3% | 60.3% | Heavy quality loss |
| IQ2_M | 20.7% | 19.0% | 19.8% | Below usable threshold |
| IQ2_XS | 5.2% | 3.4% | 4.3% | Triggered early-stop for lower bits |
Research Guidance
- Public research recommendation: Q3_K_M only
- IQ3_M is still uploadable for experiments, but quality is clearly degraded
- The rerun auto-stopped below IQ2_XS because average pass rate fell under 50%, so lower-bit quants should be considered archival artifacts rather than viable deployments
- For any user-facing scenario, prefer the release bundle instead of this research branch
For cross-family ranking and release-vs-research comparison, see results/COMPARISON.md in the workspace.
