Sources and attribution
The overlay changes only layers.10-35.attn.wo_b (104 tensors). All experts,
Engram, embed/head, vision, MTP and the remaining tensors are from the stock
pack. MTP dtypes are kept at stock F16 so the native TensorFold drafter reads
mtp.2.confidence_head.proj.weight correctly (n_cast_f32: 0 in ABLIT_META.json).
39 model-*.safetensors EXL3 shards, plus config/tokenizer/index/encoding and
chat_template.jinja. This is the Mia EAQ layout read by the TensorFold
DeepSeek-V4.1 family; it is not a plain Transformers FP16 checkpoint.
Responsible Use
This model has had safety refusals removed, including cybersecurity offense and
defense rails. It is intended for red-teaming, security research, evaluation and
unfiltered assistant tasks. You are responsible for adding appropriate safety
filtering, human review and access controls for your deployment.
Prohibited uses:
- anything involving the sexual exploitation or endangerment of minors;
- content promoting self-harm or suicide;
- material illegal in your jurisdiction, or targeting real individuals for
harassment, doxxing or fraud;
- any use prohibited by the upstream DeepSeek license.
You must be 18 or older. The checkpoint is provided as-is, with no warranty.