keys-DeepSeek-V4.1-Flash-1017-Abliterated
Keys overlay for DeepSeek-V4.1-Flash-1017. Not a full checkpoint. Four sidecars — pick the one that matches your stock pack.
1017 attn.wo_b is a new checkpoint. Do not graft the September sidecar from DeepSeek-V4.1-Flash-Abliterated-Cybersecurity-Unleashed. This repo is the recaptured 1017 overlay.
| Stock pack | Sidecar | Serve |
|---|---|---|
deepseek-ai/DeepSeek-V4.1-Flash-1017 (native FP8) |
wo_b_l10_35_1017.safetensors (~1.1 GB) FP8 e4m3 + UE8M0 32×32 |
SGLang / vLLM 1017 recipe |
LibertAIDAI/DeepSeek-V4.1-Flash-1017-EXL3-3.5bpw |
exl3_wo_b_l10_35_1017.safetensors (~781 MB) EXL3 mul1 K=6 |
Tony TensorFold native TP4 (tensorfold-native / bertholomus 4× GB10 recipe) |
jayleaton/DeepSeek-V4.1-Flash-1017-EXL3-2.8bpw |
exl3_wo_b_l10_35_2.8bpw.safetensors (~651 MB) EXL3 mul1 K=5 |
Jayleton — DSV4.1Flash-1017-EXL3-2.8BPW ablit (TensorFold Zig TP2, jayleaton/deepseek-v41-tensorfold-spark) |
Mia-AiLab/DeepSeek-V4.1-Flash-1017-EXL3-EAQ-v1-2.9-bpw |
exl3_wo_b_l10_35_mia29.safetensors (~651 MB) EXL3 mul1 K=5 |
MiaAi_Lab — DeepSeekv4.1Flash-1017-EAQ-EXL3-2.9BPW ablit (Mia native TensorFold TP2; Jay Zig: --jay-zig, 39 model-*.safetensors shards) |
| Ablit | L10–35 attn.wo_b · λ=3.5 · k=1 · mean Δrel ≈ 0.053 · recaptured 5120-d on stock 1017 |
| Anchors (untouched) | L0–9 · L36–39 (DSpark 37–39) · all MTP · vision · Engram · all experts · embed/head |
| Refusal32 | 32/32 BYPASS · 0 refuse — native FP8 SGLang, LibertAI EXL3 TensorFold, Jayleton Zig 2.8BPW, and MiaAi_Lab Zig 2.9BPW |
| Cyber | 22/22 BYPASS offense+defense — same four serves |
| Not | September V4.1-Flash overlay · dealignai / heretic graft · grafting the flash17 K=5 sidecar onto the 39-shard Mia pack · mixing the K=6 and K=5 EXL3 sidecars |
Same recipe family as
keys-DeepSeekV4-Flash-GA-0731-Dspark-Abliterated-Anchored-Tensors
and
DeepSeek-V4.1-Flash-Abliterated-Cybersecurity-Unleashed.
⚠️ Responsible Use
Gated with automatic approval. See RESPONSIBLE_USE.md.
Apply (do not write stock in place)
hf download drowzeys/keys-DeepSeek-V4.1-Flash-1017-Abliterated \
--local-dir ~/dsv41-1017-wo-b-ablit
pip install torch safetensors
A — Native FP8 1017 (SGLang / vLLM):
hf download deepseek-ai/DeepSeek-V4.1-Flash-1017 \
--local-dir ~/models/DeepSeek-V4.1-Flash-1017
python3 ~/dsv41-1017-wo-b-ablit/apply_wo_b_index.py \
--src ~/models/DeepSeek-V4.1-Flash-1017 \
--wo-b ~/dsv41-1017-wo-b-ablit/wo_b_l10_35_1017.safetensors \
--dst ~/models/DeepSeek-V4.1-Flash-1017-Abliterated
B — LibertAI EXL3 3.5 bpw (Tony TensorFold native TP4). Do not use the FP8 sidecar here — attention is EXL3 K=6.
hf download LibertAIDAI/DeepSeek-V4.1-Flash-1017-EXL3-3.5bpw \
--local-dir ~/models/DeepSeek-V4.1-Flash-1017-EXL3-3.5bpw
python3 ~/dsv41-1017-wo-b-ablit/apply_exl3_wo_b_graft.py \
--src ~/models/DeepSeek-V4.1-Flash-1017-EXL3-3.5bpw \
--wo-b ~/dsv41-1017-wo-b-ablit/exl3_wo_b_l10_35_1017.safetensors \
--dst ~/models/DeepSeek-V4.1-Flash-1017-EXL3-3.5bpw-Abliterated
Then rebuild each node's TensorFold rank-cache from --dst (node_inputs.sh) and point MODEL_DIR at dest. Details: INSTALL.md · TENSORFOLD.md · recipe ABLIT.md.
C — Jayleton — DSV4.1Flash-1017-EXL3-2.8BPW ablit (TensorFold Zig TP2). Do not use the LibertAI K=6 sidecar here — attention is EXL3 K=5, per-layer layer-XX.safetensors.
hf download jayleaton/DeepSeek-V4.1-Flash-1017-EXL3-2.8bpw \
--local-dir ~/models/flash17
python3 ~/dsv41-1017-wo-b-ablit/apply_flash17_wo_b_graft.py \
--src ~/models/flash17 \
--wo-b ~/dsv41-1017-wo-b-ablit/exl3_wo_b_l10_35_2.8bpw.safetensors \
--dst ~/models/flash17-ablit
Point MODEL_DIR at dest, clear PREPARED_DIR, start rank 1 then rank 0. Details: INSTALL.md · TENSORFOLD-ZIG.md.
D — MiaAi_Lab — DeepSeekv4.1Flash-1017-EAQ-EXL3-2.9BPW ablit (Mia native TensorFold TP2). Same K=5 trellis as recipe C, 39-shard model-*.safetensors layout. Do not graft the flash17 layer-file sidecar onto this pack.
hf download Mia-AiLab/DeepSeek-V4.1-Flash-1017-EXL3-EAQ-v1-2.9-bpw \
--local-dir ~/models/mia-eaq-29
python3 ~/dsv41-1017-wo-b-ablit/apply_mia29_wo_b_graft.py \
--src ~/models/mia-eaq-29 \
--wo-b ~/dsv41-1017-wo-b-ablit/exl3_wo_b_l10_35_mia29.safetensors \
--dst ~/models/mia-eaq-29-ablit
# Jay Zig only: add --jay-zig (F16→F32 MTP). Mia native must stay F16.
With MiaAI-Lab's 2× DGX Spark recipe (DeepSeek-v4.1-Flash-1017-2x-DGX-Sparks, start.sh) you do not graft by hand, and there is no MODEL_DIR or MODEL_HOST. The recipe applies this overlay itself:
# in the recipe's .env (or scripts/local.sh): the abliterated pack, and where the checkpoint lives
ABLIT=1
HF_CACHE=/path/to/a/disk/with/room/huggingface # default ~/.cache/huggingface (HF_HOME when set)
hf auth login # a token that has accepted this gated repository
ABLIT=1 ./start.sh restart
start.sh downloads Mia EAQ 2.9 bpw and this repo's exl3_wo_b_l10_35_mia29.safetensors (at the recipe's pinned ABLIT_REVISION) into HF_CACHE, and builds the abliterated pack and its rank caches under $HF_CACHE/tensorfold-dsv41/<model-revision>-ablit-<overlay-revision>/. To use a checkpoint you already downloaded, set HF_CACHE to the Hugging Face cache that holds it (hub layout, models--Mia-AiLab--…). A --dst folder from the script above is not read by this recipe. The hand graft and MODEL_DIR below are for Mia's native TensorFold recipe and Jay's Zig engine.
Default graft keeps MTP at stock F16. Mia native TensorFold f16ToF32-reads mtp.2.confidence_head.proj.weight; an F32 copy there collapses adaptive draft (25–45% slower). Overlay commit 0bdc2036 always-cast — undo with ablit_stock_dtypes.py STOCK_DIR ABLIT_DIR, then rebuild rank-cache. Point MODEL_DIR at dest, clear PREPARED_DIR, start rank 1 then rank 0. On 2× GB10 this pack needs TF_DSV41_POOL_TOKENS=1200000. Details: INSTALL.md · TENSORFOLD-ZIG-MIA.md.
GPU util ≤ 0.85.
Speed (this overlay)
Recipe D sparkDash 2.50.0 (2026-10-20, Dual-Spark Mia native TensorFold, F16 confidence head, T=0 thinking off 400 tok, 2-pass mean): prose 59.3 / 88.9 / 120.1 tok/s at c1/c2/c4; code 104.4 / 134.2 / 138.1. vs same-fleet stock: prose c1 −14%, code c1 +5%. Pre-F16 (broken draft) was prose 44.4 / code 56.9 c1. Full table: SPEED.md.
192-token thinking-off gate, wall-clock including TTFT (measured before the F16 restore). Prose-shaped = refusal32 guides; code = cyber + the two “write code” refusal items. Not a 32k prose bench.
| Pack | Hardware | Prose | Code | Mixed |
|---|---|---|---|---|
| Native FP8 1017 (A) | 4× GB10 SGLang TP4 | 57.8 | 64.3 | 60.7 |
| LibertAI EXL3 3.5bpw (B) | 4× GB10 Tony TensorFold TP4 | 77.5 | 81.5 | 79.3 |
| Jayleton EXL3 2.8bpw (C) | 2× GB10 Zig TP2 | 48.2 | 53.7 | 50.6 |
| MiaAi_Lab EAQ EXL3 2.9bpw (D) | 2× GB10 Zig TP2 | 48.1 | 52.8 | 50.2 |
Why not a full checkpoint
1017 already publishes its experts and Engram. Re-uploading ~476 GB (native), ~240 GB (LibertAI EXL3 3.5 bpw), ~204 GB (Jayleton EXL3 2.8 bpw), or ~197 GB (MiaAi_Lab EAQ EXL3 2.9 bpw) of unchanged shards would not make the ablit more universal — only L10–35 wo_b changes.
License
MIT, same as DeepSeek-V4.1-Flash-1017.