The upstream abliteration is an anchored-tensor edit: it modifies only
26 attention output projections (layers.10-35.attn.wo_b.weight, lambda 3.5),
leaving all experts, the MTP/DSpark drafter, and vision untouched. Those 26
tensors are BF16 natives in our pack, so this variant is the MixedK pack with
exactly those 26 tensors swapped for their abliterated versions — no
re-encoding. Every expert is byte-identical to the non-abliterated MixedK.
Notes
~95 GiB, fits one DGX Spark. Format, serving, and limitations are identical
to the non-ablit MixedK
(mixed-format, not boot-tested, K2 double-quant).
Uncensored: safety refusals removed upstream. For red-teaming / security
research / evaluation. You bear deployment responsibility.
DeepSeek license (other) + the upstream Responsible Use terms.
