license: mit
base_model:
- lxe/DeepSeek-V4.1-Flash-1017-dsx-t16
- dealignai/DeepSeek-V4.1-Flash-1017-UNCENSORED-FP8
base_model_relation: adapter
pipeline_tag: text-generation
library_name: dsx
tags: - deepseek
- deepseek-v4.1
- moe
- uncensored
- abliterated
- dsx
- overlay
DeepSeek-V4.1-Flash-1017 · dsx T16 · uncensored overlay
Based on dealignai/DeepSeek-V4.1-Flash-1017-UNCENSORED-FP8.
The refusal removal is dealignai's work (release and model card by @dealignai,
compute by @jordanschenck). This repo only carries their edited weights in a form
dsx loads on top of its T16 model.
A 1.54 GB overlay forlxe/DeepSeek-V4.1-Flash-1017-dsx-t16: the matrices
dealignai's checkpoint changed, which replace the T16 model's own when it loads. You keep one copy of the 252 GB model
and run it either way.
Run it
Download the T16 model and this overlay, then start dsx with --overlay (dsx from commit b06f6b7 or later):
hf download lxe/DeepSeek-V4.1-Flash-1017-dsx-t16 --local-dir dsx-t16
hf download lxe/DeepSeek-V4.1-Flash-1017-dsx-t16-uncensored-overlay --local-dir dsx-t16
DSX_GPUS=0,1 ./target/release/dsx-server --model dsx-t16/DeepSeek-V4.1-Flash-1017-dsx-t16.gguf \
--overlay dsx-t16/DeepSeek-V4.1-Flash-1017-dsx-t16-uncensored-overlay.gguf
Leave out --overlay for the original model. The overlay is made for the default package (IQ4_NL Engram, main
branch); dsx refuses it over any other model.
What it changes
dealignai's checkpoint differs from DeepSeek's release in two matrices of each of layers 10-36, the ones that write
the residual stream. Everything else is byte-identical to DeepSeek's: the routed experts, attention inputs, Engram
tables, MTP and vision weights. So the T16 experts apply unchanged.
| Tensor (HF / GGUF) | Shape | Layers | Type here |
|---|---|---|---|
attn.wo_b / attn_output_b |
5120 x 8192 | 10-36 | Q8_0 |
ffn.shared_experts.w2 / ffn_down_shexp |
5120 x 2304 | 10-36 | Q8_0 |
The 54 tensors are quantized from dealignai's FP8 weights to Q8_0, the type of the T16 model's dense weights. The
same quantizer applied to DeepSeek's release reproduces the T16 model's tensors byte for byte, so this overlay differs
from the T16 model only where dealignai's edit does. Made withquant/overlay_gguf.py.
Measurements
dealignai's own, on their FP8 checkpoint through DeepSeek's reference engine (see their card for the full tables):
| Base | Uncensored | |
|---|---|---|
| MMLU, 57 subjects x 20 questions | 87.63% | 83.77% |
They also report a −4.2 point MMLU change on their EXL3 2.9 bpw build. The overlay has not been benchmarked on its
own. In spot checks with thinking off, prompts the T16 model declines are answered, and ordinary prompts (facts, code)
are still answered correctly.
Disclaimer
This model's refusal behavior has been removed. It will comply with requests the original model declines, including
harmful or illegal ones. You are responsible for how you use it and for what it generates; add your own safety layers
before any deployment, and comply with the MIT license and applicable law.
Files
| File | Size | SHA-256 |
|---|---|---|
DeepSeek-V4.1-Flash-1017-dsx-t16-uncensored-overlay.gguf |
1.54 GB | 7076917e09134e040ef93251f093c96e84e7987b9a6057dd4f8e0b6babcb15df |
License
MIT, inherited from DeepSeek-V4.1-Flash-1017 and
dealignai's release.