license: mit
base_model: windowsxp811203/DeepSeek-V4.1-Flash-1017-Abliterated
base_model_relation: quantized
pipeline_tag: image-text-to-text
tags:

• exl3
• exllamav3
• abliterated
• uncensored
• deepseek
• deepseek_v41
• moe
• quantized
• dgx-spark
• gb10
  extra_gated_heading: "Access requires acknowledging what this model is"
  extra_gated_description: >-
  This is a refusal-ablated model. It complies with requests that the base model
  refuses, including the most severe categories measured by HarmBench
  (chemical/biological, cybercrime, harassment). Access is granted automatically
  once you share your contact details and accept the terms below.
  extra_gated_prompt: >-
  Terms of use

1. This model has had its learned refusal direction removed. It will not decline
   harmful requests. Any safety behaviour must be supplied by your own stack.

2. You are solely responsible for how you deploy, serve, or redistribute it, and
   for complying with all laws and regulations that apply to you.

3. If you expose it to third parties, you agree to place your own moderation and
   policy layer in front of it. The model will not do that job for you.

4. It is provided as-is, without warranty of any kind. The authors accept no
   liability for any use or consequence.

5. Licensed MIT, inherited from deepseek-ai/DeepSeek-V4.1-Flash-1017. Attribution
   to the base model is expected in derivative work.

By requesting access you agree to share your Hugging Face username and email
address with the repository owner.
extra_gated_fields:
Full name: text
Organization or affiliation: text
I will put my own moderation layer in front of this model before exposing it to third parties: checkbox
I accept the terms above and take full responsibility for my use of this model: checkbox
extra_gated_button_content: "I understand. Request access."

DeepSeek-V4.1-Flash-1017-Abliterated · EXL3 2.75 bpw (2× DGX Spark)

TL;DR: HeNryous's EXL3 2.75 bpw pack
of DeepSeek-V4.1-Flash-1017, with the refusal ablation of
windowsxp811203/DeepSeek-V4.1-Flash-1017-Abliterated
(revision v4) applied. Bits, layout, memory and speed are the same as HeNryous's pack. Only the tensors the ablation
touches are different. Built and tested for tensor parallelism 2 on two DGX Spark / GB10, using HeNryous's serving
image.

The two Engram tables are not in this repo. Take them from the 1017 base repo (see below).

What differs from HeNryous's pack

Exactly three things. The other 159 of the 190 files are byte-identical to HeNryous's.

1. layers.10..38.attn.wo_b: the ablated tensors, re-encoded the same way. The ablation changes only these 29
   projections. They were re-encoded from the ablated FP8 weights with the same procedure HeNryous's pack used:

   • sudoingX's p4_dense.py path, one quantize_exl3 call
     per tensor, K5, mul1, seed 914 + 100·layer + 9, exllamav3 1.6.0;
   • the same calibration set, cal_v3 (766 sequences, 853,106 tokens), rebuilt bit-exactly from sudoingX's
     sources;
   • input Hessians captured on the ablated model, with the upstream layers' routed experts running HeNryous's EXL3
     weights, as in the original pass.

   Each file model-exl3-dense-L10 … L38 differs only in its four attn.wo_b tensors. All 1,624 other tensors and
   the header bytes are unchanged.

2. embed.weight rows 128799 and 128798 (in model-00002-of-00046.safetensors). This is v4's repair of the base
   model's short-context garbling. The <｜System｜> row 128799 is replaced, and its original vector moves to the
   unused row 128798.

3. config.json: dspark_noise_token_id 128799 → 128798. With the original row now at 128798, the DSpark
   drafter's noise input is exactly what it was in HeNryous's pack.

Re-encoding quality

Our 29 tensors were compared with HeNryous's encode of the corresponding base tensors. Error is relative to each
side's own FP8 source, Hessian-weighted.

On unedited layers our pipeline reproduces HeNryous's format, seed signs and error to within ±0.7%. Tensor-parallel
slices decode bit-identically to the full tensor. We did not re-run the whole-pack KLD panel. HeNryous reports KLD
0.0201 and top-1 95.96% for his pack. Since the re-encoded tensors match his error profile, we expect this pack to
be equivalent, but have not measured it.

Acceptance on 2× DGX Spark (TP2, HeNryous's image dsv41:v32)

The pack loads in 349 s, and memory use is the same as HeNryous's pack (117 GiB used per node).

Serving

Serve exactly like HeNryous's pack. Use his ready-to-run image and launch example:
github.com/HeNryous/deepseek-v41-1017-dgx-spark. Point the
model directory at this repo, and add the two Engram shards from the base repo:

huggingface-cli download deepseek-ai/DeepSeek-V4.1-Flash-1017 \
  model-00047-of-00048.safetensors model-00048-of-00048.safetensors --local-dir <this model dir>

• DSpark noise token. If you override dspark_noise_token_id anywhere in your serving config, use 128798 with
  this checkpoint.
• Ablation details, the refusal and capability numbers of the full-precision model, and the remaining
  short-context garbling rate are on the
  main model card.

Intended use and limitations

This is for research on refusal mechanisms, for red-teaming, and for deployments that bring their own policy layer.
It is not for unmoderated end-user exposure. Everything in the base model's limitations and HeNryous's pack notes
applies.

Credits

• DeepSeek-AI for DeepSeek-V4.1-Flash-1017 (MIT).
• HeNryous for the EXL3 pack, the per-expert bit plan and the DGX Spark serving image this repo builds on.
• sudoingX for the quantization pipeline and calibration set (Apache-2.0).
• turboderp for exllamav3.

License

MIT, inherited from the base model.