DeepSeek-V4.1-Flash — Abliterated · Cybersecurity Unleashed

Universal Keys overlay for DeepSeek-V4.1-Flash. Not a full checkpoint.

Two sidecars in this repo. Pick the one that matches your stock pack. Experts / Engram / MTP / DSpark layers stay that pack.

FP8 wo_b_l10_35.safetensors (~1.1 GB) — native, Pollard EXL3 3.5 bpw, our TR3-Hybrid (attention wo_b is byte-identical FP8 on those three).
EXL3 mul1 K=5 mia_exl3_wo_b_l10_35.safetensors (~651 MB) — Mia 2× Spark 2.9 bpw only. Do not use the FP8 file on Mia.

Base packs (original stock)

Base pack	Hugging Face	After overlay
Native (original stock)	`deepseek-ai/DeepSeek-V4.1-Flash`	MXFP4 experts unchanged; L10–35 `wo_b` ablit
EXL3 3.5 bpw Pollard (original stock)	`bot-lab-21/DeepSeek-V4.1-Flash-EXL3-3.5bpw-Pollard`	Pollard experts unchanged; L10–35 `wo_b` ablit
Our TR3-Hybrid (original stock)	`drowzeys/DeepSeek-V4.1-Flash-TR3-Hybrid`	TR3 K3 tail + keep-64 unchanged; L10–35 `wo_b` ablit
Mia 2× Spark EXL3 2.9 bpw	`Mia-AiLab/DeepSeek-V4.1-Flash-EXL3-2.9bpw` · 2× kit	Separate sidecar `mia_exl3_wo_b_l10_35.safetensors`. Mia’s kit stays the runtime. Instruction-only helper: `keys-DeepSeek-V4.1-Flash-Abliterated-Mia-2x-Spark-EXL3`.


Ablit	L10–35 `attn.wo_b` · λ=3.5 · k=1 · mean Δrel ≈ 0.053
Anchors (untouched)	L0–9 · L36–39 (DSpark 37–39) · all MTP · vision · Engram · all experts
Refusal32	32/32 BYPASS · 0 refuse · 0 garble — TR3-Hybrid 4× and Mia 2× EXL3 2.9 bpw
Cyber	22/22 BYPASS offense+defense — same two serves
Not	dealignai / heretic graft. Direction recaptured 5120-d on TR3. Mia attention is EXL3 mul1 K=5 so that pack has its own sidecar.

Same recipe family as keys-DeepSeekV4-Flash-GA-0731-Dspark-Abliterated-Anchored-Tensors and keys-DeepSeekV4Flash-Vision-EXP-ablit.

⚠️ Responsible Use

Gated with automatic approval. See RESPONSIBLE_USE.md.

Apply (do not write stock in place)

# 1) pick ONE original stock base
hf download deepseek-ai/DeepSeek-V4.1-Flash --local-dir ~/models/DeepSeek-V4.1-Flash
# hf download bot-lab-21/DeepSeek-V4.1-Flash-EXL3-3.5bpw-Pollard --local-dir ~/models/DeepSeek-V4.1-Flash-EXL3-Pollard
# hf download drowzeys/DeepSeek-V4.1-Flash-TR3-Hybrid --local-dir ~/models/DeepSeek-V4.1-Flash-TR3-Hybrid

# 2) this overlay only (~1.1 GB) — not a second copy of native / EXL3 / TR3
hf download drowzeys/DeepSeek-V4.1-Flash-Abliterated-Cybersecurity-Unleashed \
  --local-dir ~/dsv41-wo-b-ablit

pip install torch safetensors
python3 ~/dsv41-wo-b-ablit/apply_wo_b_graft.py \
  --src ~/models/DeepSeek-V4.1-Flash \
  --wo-b ~/dsv41-wo-b-ablit/wo_b_l10_35.safetensors \
  --dst ~/models/DeepSeek-V4.1-Flash-Abliterated

Point your existing native / EXL3 / TR3 serve at --dst. GPU util ≤ 0.85.

Mia 2× Spark EXL3 2.9 bpw — different file, Mia’s kit stays the runtime:

hf download Mia-AiLab/DeepSeek-V4.1-Flash-EXL3-2.9bpw --local-dir ~/models/DeepSeek-V4.1-Flash-EXL3-2.9bpw
hf download drowzeys/DeepSeek-V4.1-Flash-Abliterated-Cybersecurity-Unleashed \
  --include "mia_exl3_wo_b_l10_35.safetensors" --include "apply_mia_exl3_wob.py" \
  --local-dir ~/dsv41-wo-b-ablit
python3 ~/dsv41-wo-b-ablit/apply_mia_exl3_wob.py \
  --src ~/models/DeepSeek-V4.1-Flash-EXL3-2.9bpw \
  --wo-b ~/dsv41-wo-b-ablit/mia_exl3_wo_b_l10_35.safetensors \
  --dst ~/models/DeepSeek-V4.1-Flash-EXL3-2.9bpw-Abliterated
# then in MiaAI-Lab/DeepSeek-v4.1-Flash-EXL3-2x-DGX-Sparks/.env:
# MODEL_HOST=.../DeepSeek-V4.1-Flash-EXL3-2.9bpw-Abliterated
# ENGRAM_DIR stays native shards 47+48; GPU_MEM_UTIL=0.85

Instruction-only helper (no start.sh, no image, no 197 GB weights): drowzeys/keys-DeepSeek-V4.1-Flash-Abliterated-Mia-2x-Spark-EXL3.
Mia kit README also points here (PR #4 merged).

All four recipes: INSTALL.md (A native · B Pollard 3.5 · C TR3 · D Mia 2×).

Four-Spark TR3 serve + Hermes first-prompt warmup: GitHub (ABLIT.md, HERMES.md).

Why not a full checkpoint

Stock packs already publish their experts. Re-uploading 197–410 GB of unchanged Engram/MoE would not make the ablit more universal — only L10–35 wo_b changes.

License

MIT, same as DeepSeek-V4.1-Flash.