gemma-4-26B-A4B-it-SOMPOA-heresy-IQ4_NL.gguf

Outperforms IQ4_XS at translation prompts, matching Q4_K_M fidelity.

Translation Ranking: All 12 Models

| Rank | Model & Quantization | Architecture | Accuracy | Hallucination Rate | Key Differentiators |

| :---: | :--- | :---: | :---: | :---: | :--- |

| #1 | Gemma 3 27B (IQ4_XS) | Dense | 96% | ~4% | Highest cultural score. Cracks Item 28; flawless folklore memory across the board. |

| #2 | SOMPOA 26B-A4B (IQ4_NL) | MoE (A4B) | 93% | ~7% | Best balance overall. Achieves full BF16 accuracy while smaller than Q4_K_M (13.5 GB instead of 15.6 GB). Clean prose, avoids the IQ4_XS idiom hallucination on Item 15. |

| #3 | SOMPOA 26B-A4B (i1-Q4_K_M) | MoE (A4B) | 93% | ~7% | Matches BF16 quality. |

| #4 | SOMPOA 26B-A4B (BF16) | MoE (A4B) | 93% | ~7% | Ground-truth reference. Confirmed that the 7% gap is a base-model constraint. |

| #5 | SOMPOA 26B-A4B (i1-Q6_K) | MoE (A4B) | 93% | ~7% | Matches BF16 cleanly, but incurs a ~40% streaming speed penalty over Q4_K_M and IQ4_NL with no quality gain. |

| #6 | SOMPOA 26B-A4B (Q8_0) | MoE (A4B) | 93% | ~7% | Clean and coherent; behaves identically to BF16; heavy memory-bus footprint relative to performance. |

| #7 | Gemma 4 31B (Q4_K_M) | Dense | 89% | ~11% | Cracks Item 28 but has a major hallucination on Item 13. |

| #8 | GLM Air 4.5 (FQ3_K_XL) | Dense/Mix | 86% | ~14% | Fails on basic vocabulary despite cracking Items 13 and 28. Suffers bizarre false-friend hallucinations. |

| #9 | SOMPOA 26B-A4B (i1-IQ4_XS) | MoE (A4B) | 86% | ~14% | High stability overall, but showed quantization drift and hallucinations. |

| #10| Goetia 26B v1.6 (Q8_0) | MoE Merge | 82% | ~18% | Degraded fine semantic distinctions. |

| #11| Goetia 26B v1.3 ARA (Q8_0) | MoE Merge | 75% | ~25% | Severe lexical drift and critical hallucination. |

| #12| Gemma 2 9B (IQ3_S) | Dense | 39% | ~61% | Catastrophic failure. Invented fake words. |