Ling-3.1-flash-1019 (abliterated) — GGUF

GGUF quants of mrfakename/Ling-3.1-flash-1019-abliterated, which is inclusionAI/Ling-3.1-flash-1019 with a rank-4 refusal subspace projected out of layers 16–71. The model card explains what was measured, what was edited and how much is left — including the three separate places this model keeps its refusals, and the state-aligned reframes that a refusal classifier scores as compliance.

560B total parameters, 25B active. Needs llama.cpp with bailingmoe3 support.

Quants

Every quant is split into shards; llama.cpp loads the whole set when you point it at the first file.

hf download mrfakename/Ling-3.1-flash-1019-abliterated-GGUF --include 'Q4_K_M/*' --local-dir .
llama-server -m Q4_K_M/Ling-3.1-flash-1019-abliterated-Q4_K_M-00001-of-*.gguf

The K-quants are requantized from the Q8_0 rather than from BF16 — a BF16 GGUF of a 560B model is 1.1 TB and does not fit on a box next to its own output. quantize-logs/ has the per-tensor logs; in each K-quant, 10 of 1626 tensors fall back to a neighbouring type because their shape does not divide the block size, which is normal.

LoRA form

abliteration-lora.gguf is the same edit as a rank-4 LoRA, so it can be applied to anyone's quant of the stock release instead of downloading this one:

llama-server -m Ling-3.1-flash-1019-Q8_0-00001-of-00015.gguf --lora abliteration-lora.gguf

Patch llama.cpp first. bailingmoe3 writes the attention output with a raw ggml_mul_mat instead of build_lora_mm, so llama.cpp quietly ignores any LoRA on attn_output — which is 56 of this adapter's 107 tensors, and the half with the larger strengths. Without the patch the adapter still loads, still reports as applied, and does roughly half of the edit. bailingmoe3-lora-attn.patch here is the three-line fix (two attention kinds plus the MTP block); it applies to current master as well as to the pinned commit this was built against.

git apply bailingmoe3-lora-attn.patch && cmake --build build -j

The adapter is exact for the tensors it touches (it is -s·Q(QᵀW) built from the original weights), but applying a LoRA to an already-quantized tensor is not bit-identical to editing and then quantizing, so the patched checkpoint in the model repo is the reference.

This model will comply with harmful requests. You are responsible for how you use it.