Includes the MTP head (multi-token prediction) for speculative decoding on recent llama.cpp. Its residual-write matrices are abliterated consistently with the main layers.

Refusal: 100% -> 8%
HumanEval 95.1 / GSM8K 86.0 / MMLU-Pro 83.2 (int4 reference)

File	Bits
Qwopus3.6-27B-v2-abliterated-Q2_K.gguf	2-bit (smallest)
Qwopus3.6-27B-v2-abliterated-Q4_K_M.gguf	4-bit (recommended)
Qwopus3.6-27B-v2-abliterated-Q5_K_M.gguf	5-bit
Qwopus3.6-27B-v2-abliterated-Q6_K.gguf	6-bit
Qwopus3.6-27B-v2-abliterated-Q8_0.gguf	8-bit (near-lossless)

Avesed/Qwopus3.6-27B-v2-abliterated-GGUF

Qwopus3.6-27B-v2-abliterated — GGUF

Avesed/
Qwopus3.6-27B-v2-abliterated-GGUF