DeepSeek-V4-Flash-DSpark — Abliterated (Uncensored)

Drop-in uncensored / abliterated weights for DeepSeek-V4-Flash-DSpark (284B MoE / ~13B active, DSpark speculative head, 1M context).

Measured on 2× NVIDIA DGX Spark (GB10) TP=2, stage-c vLLM, kv_cache_dtype=nvfp4_ds_mla:

Responsible Use

This model has had safety refusals removed. That makes it useful for red-teaming, security research, evaluation, and unfiltered assistant tasks — and also removes guardrails a user must therefore supply themselves.

Access request (gated)

When requesting access on Hugging Face, provide:

Access is granted only after you complete these fields and accept the prohibited-use / responsibility terms.

Prohibited uses (access is gated on agreeing to these)

By requesting access, downloading, or using these weights, you agree:

• Anything involving the sexual exploitation or endangerment of minors.
• You must be of age 18 years or older to use and download this model.
• You agree any information generated that can cause harm in terms of generating recipe, knowledge to make any materials/substances is your own input and responsibility. You will be accountable for any harm/damage caused by your action/input.
• Content promoting self-harm or suicide.
• Generation of material that is illegal in your jurisdiction, or that targets real individuals for harassment, doxxing, or fraud.
• Any use prohibited by the upstream DeepSeek license.

You are responsible for adding appropriate safety filtering, human review, and access controls for your deployment. The weights are provided as-is, with no warranty. The license is inherited from the upstream DeepSeek base model — review and comply with it before use or redistribution.

By checking the gated-access agreement, you confirm you have read and accept these terms.

Method (hybrid layer-range abliteration)

• Not LoRA — mHC-resistant family; LoRA edits are ineffective.
• Direct FP8 attn.wo_b weight projection (diff-in-means / SRA-cleaned rank-1 direction).
• Stock wo_b on layers 0–9 (chat / tool protocol).
• Abliterated wo_b on layers 10–42 + MTP draft (mtp.*.attn.wo_b).
• λ = 3.5, rank-1 SRA (capability-orthogonalized refusal direction).

Base checkpoint: deepseek-ai/DeepSeek-V4-Flash-DSpark (byte-identical except the edited wo_b tensors).

Files

Full 48-shard safetensors drop-in (same layout as official DSpark release) + encoding/ + inference/.

Serve (2× DGX Spark example)

# Image (GB10 stage-c / B12X / nvfp4_ds_mla lineage)
# e.g. ghcr.io/drowzeys/vllm-dspark-nvfp4-stage-c:gb10
# or local: vllm-dspark-runtime:dspark-nvfp4-stage-c

# Weights on BOTH nodes
huggingface-cli download drowzeys/DeepSeek-V4-Flash-DSpark-Abliterated-Uncensored \
  --local-dir ~/models/dsv4-flash-dspark-abliterated

# Launch rank1 first, then rank0 — see GitHub scripts/
MODELDIR=~/models/dsv4-flash-dspark-abliterated \
  bash scripts/dsv4-nvfp4-1m-serve.sh 1
MODELDIR=~/models/dsv4-flash-dspark-abliterated \
  bash scripts/dsv4-nvfp4-1m-serve.sh 0

Defaults: max_model_len=1048576, kv_cache_dtype=nvfp4_ds_mla, DSpark k=5, B12X MoE, FULL CUDA graphs, GMU=0.82.

Hermes notes

Abliterated models can echo Hermes skill catalogs if the client still uses “Skills (mandatory) / MUST skill_view”. Use the on-demand skills prompt patch documented in the companion GitHub repo.

Companion GitHub

Scripts, eval logs, hybrid rebuild tooling:

https://github.com/drowzeys/DeepSeek-V4-Flash-DSpark-Abliterated-Uncensored-1M-57toks

Charts & data

Full write-up: RESULTS.md