TaQuants/Gemma-4-12B-OBLITERATED-TaQuants-GGUF

Gemma-4-12B-OBLITERATED TaQuants

The Repository Technical Report

The Tema_Q development team, team zenei, has developed a new importance matrix method called TaQuants (Tensor-aware Adaptive Quantization). This model is a TaQuants version of OBLITERATUS/Gemma-4-12B-OBLITERATED created with TaQuants v3.0.

The model size and performance are as follows: | | Size (GB) | Knowledge score | | --- | --- | --- | | Q4KM | 7.38 | 31.97 | | TaIQ3S | 5.48 | 31.88 | | TaIQ2M | | 9.09 |

Vulnerabilities to quantization were identified in the model. Q4KM caused partial collapse, while TaIQ2M exhibited complete collapse. TaIQ3S performs inference comparable to Q4KM but avoids the partial collapse observed with Q4KM.