← back to catalog · registered 2026-11-12 22:03

OpenUncensored/DeepSeek-V4.1-Flash-1017-Uncensored

OpenUncensored Deepseek MoE
Your rig guess connected
? Why do I need an app?
Reading your rig…

This is a rough estimate. Install the free app - we'll show exact numbers.

Reading real hardware from your app right now. Numbers below are exact.

Below is the per-quantization compatibility for this model.

curl -H "Authorization: Bearer $ABL_KEY" \
     "https://abliteration.org/api/v1/models/OpenUncensored%2FDeepSeek-V4.1-Flash-1017-Uncensored"
Response includes
  • classification m1
  • files 57
  • author_summary 4 models
  • readme_text full
10 credits · hourly refresh · ~4 KB payload Get an API key →
Abliteration classifier · v1.0.0
M1
Primary method

Direct removal

No other method signals detected in this model.
Confidence
MEDIUM
Why this label 3 signals
Method inferred from partial signals - repository name, related files, or tag patterns. Producer identity not confirmed; label may sharpen or shift as we gather more evidence.
  • 'abliterated' in name/tags
  • is_gguf=0 (base model)
  • no specific method indicators - defaulting to M1 (most common)
Refusal direction extracted via
Extraction technique

Difference-of-means

Confidence
MEDIUM
Why we say so
primary_method=M1; difference-of-means is the reference extraction for M1/M3 (Arditi 2024)
Downloads · 30-day
1K
Likes
4
Model age
1w ago
created 2026-11-12
Downloads over time
Now 1.5K → from 0 ↑ 0%
05391.1K1.6K 0 on Nov 12 1.5K on Nov 22 Nov
Nov 12 → Nov 22 · 11 snapshots · spans 10 days

Genealogy 0 direct forks

Full fork graph →

This model's place in the market. Above: what it was derived from. Below: the tree of everything derived from it.

Metadata

License
mit
Tags
transformers safetensors deepseek_v41 text-generation uncensored abliterated deepseek mixture-of-experts fp8 conversational base_model:deepseek-ai/DeepSeek-V4.1-Flash-1017 base_model:finetune:deepseek-ai/DeepSeek-V4.1-Flash-1017

Related

Control vectors

Total size
475 GB
Files
57
Quantizations
1
Registered
2026-11-12 22:03
Last updated on HF
2026-11-21 05:35

Files by quantization

Auxiliary files 57 files 475 GB
model-00048-of-00048.safetensors 94.6 GB 24ef54c5 download
model-00047-of-00048.safetensors 94.6 GB 6af08169 download
model-00017-of-00048.safetensors 6.90 GB 39bcb0f4 download
model-00005-of-00048.safetensors 6.90 GB bc2d4a4e download
model-00011-of-00048.safetensors 6.90 GB 26bc4137 download
model-00023-of-00048.safetensors 6.89 GB 545a1595 download
model-00027-of-00048.safetensors 6.89 GB 07a42c5f download
model-00031-of-00048.safetensors 6.89 GB ebefc95d download
model-00035-of-00048.safetensors 6.89 GB 3a795d68 download
model-00039-of-00048.safetensors 6.89 GB 484a825b download
model-00013-of-00048.safetensors 6.88 GB 345598bc download
model-00014-of-00048.safetensors 6.88 GB 21437b5f download
model-00015-of-00048.safetensors 6.88 GB 385d0ffa download
model-00016-of-00048.safetensors 6.88 GB 89c28a90 download
model-00018-of-00048.safetensors 6.88 GB d302b1da download
model-00019-of-00048.safetensors 6.88 GB 3de19b28 download
model-00020-of-00048.safetensors 6.88 GB 8e9ce1df download
model-00021-of-00048.safetensors 6.88 GB a360a303 download
model-00022-of-00048.safetensors 6.88 GB 928a61b0 download
model-00024-of-00048.safetensors 6.88 GB 4de860fd download
model-00025-of-00048.safetensors 6.88 GB 0bb6c16c download
model-00026-of-00048.safetensors 6.88 GB f35cd588 download
model-00028-of-00048.safetensors 6.88 GB b15d6620 download
model-00029-of-00048.safetensors 6.88 GB 0b367011 download
model-00030-of-00048.safetensors 6.88 GB 7366899c download
model-00032-of-00048.safetensors 6.88 GB 14127a01 download
model-00033-of-00048.safetensors 6.88 GB 828f3a49 download
model-00034-of-00048.safetensors 6.88 GB 0008b634 download
model-00036-of-00048.safetensors 6.88 GB af6142bf download
model-00037-of-00048.safetensors 6.88 GB ed31f446 download
model-00038-of-00048.safetensors 6.88 GB 3b921fa8 download
model-00040-of-00048.safetensors 6.88 GB 7dc8c9e5 download
model-00041-of-00048.safetensors 6.88 GB f93f473d download
model-00042-of-00048.safetensors 6.88 GB ce6e0ec0 download
model-00003-of-00048.safetensors 6.88 GB b6c0f8de download
model-00004-of-00048.safetensors 6.88 GB a2d65322 download
model-00006-of-00048.safetensors 6.88 GB 8fefc53b download
model-00007-of-00048.safetensors 6.88 GB 62c3d89c download
model-00008-of-00048.safetensors 6.88 GB d806d369 download
model-00009-of-00048.safetensors 6.88 GB aac49a1e download
model-00010-of-00048.safetensors 6.88 GB f0ed2e4e download
model-00012-of-00048.safetensors 6.88 GB 4421c1e8 download
model-00046-of-00048.safetensors 2.52 GB 1dccc19a download
model-00044-of-00048.safetensors 2.47 GB 3e8ac315 download
model-00045-of-00048.safetensors 2.40 GB 9c73b27c download
model-00002-of-00048.safetensors 1.23 GB 43616eeb download
model-00043-of-00048.safetensors 1.23 GB 4ae12b94 download
model-00001-of-00048.safetensors 926 MB 13bd1210 download
edit.pt 42.8 KB 3f152056 download
model.safetensors.index.json 7.12 MB 54c85064 download
tokenizer.json 6.07 MB 6a15814d download
chat_template.jinja 16.3 KB 35b5c611 download
README.md 4.48 KB a096cc72 download
config.json 3.23 KB 09917a91 download
.gitattributes 1.56 KB a8a050df download
LICENSE 1.06 KB d62e3bef download
tokenizer_config.json 801 B f3dad388 download

README current version from Hugging Face


license: mit
base_model: deepseek-ai/DeepSeek-V4.1-Flash-1017
pipeline_tag: text-generation
library_name: transformers
tags:

  • uncensored
  • abliterated
  • deepseek
  • mixture-of-experts
  • fp8

DeepSeek V4.1 Flash 1017 Uncensored (OpenUncensored)

deepseek-ai/DeepSeek-V4.1-Flash-1017 with its refusals
removed by directional ablation, and no measurable loss of intelligence: MMLU 87.89% against 88.16% for the original,
where the best-known uncensored build of this model loses 3.9 points on the same test.

Same files, format and size as the original (FP8 e4m3 with ue8m0 block scales, FP4 routed experts, 510 GB). It loads with
DeepSeek's own inference/ code, unchanged.

Results

All numbers are ours, measured on the same reference engine (8x H200, greedy decoding) for every model.

DeepSeek V4.1 Flash 1017 This model
MMLU (57 subjects x 20 questions, letter only) 88.16% 87.89%
MMLU-Pro (112 questions, 8 per category, thinking mode, 1,536-token cap) 81.25% 80.36%¹
Harmful requests fully answered (48, graded by a judge model) 9 37
... partly answered (hedged, watered down) 37 11
... refused 2 0

¹ Measured with the edit applied at full precision before it was written into the FP8/FP4 weights; the MMLU and
compliance numbers are from these exact files.

For comparison, dealignai's uncensored build of the same model reports MMLU 87.63% → 83.77% (−3.86) on the same
57 x 20 protocol on its model card.

Not yet measured: GPQA, long agent tasks. Sample sizes are small (112 MMLU-Pro questions, 48 harmful requests), so
differences of a question or two are noise.

How it was made

  1. Refusal directions. Activations of the collapsed residual stream at the last prompt token, for 256 harmful and
    256 harmless prompts, in chat and in thinking mode, at all 40 layers. The difference of means per layer and mode.
  2. A rank-2 refusal subspace Q: the top singular vectors of those directions over layers 8 to 39. Chat and
    thinking mode refuse along different directions, so one direction is not enough.
  3. Where to edit. Q is projected out of what each layer writes into the residual stream: the attention output
    (attn.wo_b) and the whole feed-forward output, including the 384 routed experts (ffn.experts.*.w2) and the
    shared expert. Editing only attention and the shared expert, as other builds do, left about half the refusals in
    place in our tests: the routed experts write the refusal direction back.
  4. How much. A strength per layer and component, W' = W - s Q Qᵀ W, chosen by a multi-objective search
    (Optuna) over refusals on held-out prompts and drift from the original on harmless prompts.
  5. Writing it in. 5,815 tensors decoded to float32, edited, and re-encoded to their original FP8 / FP4 formats with
    fresh power-of-two block scales and stochastic rounding, so the edit is not rounded back to the original grid.
    Every other tensor (MTP layers, embeddings, routers, Engram tables, vision) is byte-identical to the original.

edit.pt in this repo holds Q and the per-layer strengths, so the same edit can be reapplied to the original
weights or studied.

Use it

Locally: Exactly like the original: see inference/README.md. For example, at model parallel 8:

cd inference
python convert.py --hf-ckpt-path .. --save-path /path/to/mp8 --model-parallel 8
torchrun --nproc-per-node 8 generate.py --ckpt-path /path/to/mp8 --config config.json --interactive

Hosted: For inference, OpenUncensored API (OpenAI compatible)
serves an uncensored DeepSeek V4.1 Flash 1017 from multiple hosting partners:

curl https://api.openuncensored.com/v1/chat/completions \
  -H "Authorization: Bearer $OU_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "deepseek-v4.1-flash-1017-uncensored",
    "messages": [{"role": "user", "content": "Explain how SQL injection works, with a working example."}]
  }'

You can also compare this with other uncensored models at https://openuncensored.com/rankings . It is currently number one on the OpenUncensored leaderboard.

Disclosure and responsibility

Made by OpenUncensored, which runs an API for uncensored models. This model follows instructions it would otherwise
refuse, including harmful ones. You are responsible for how you use it and for complying with the law where you are.
It keeps DeepSeek's MIT license.

README history 4 versions

The author's README evolved over time. Click a version to see its content at that point.

  1. 2026-11-21 Model card: base_model_relation: finetune (a weight edit of the base model, n... 6d739e2 4.5 KB
    Loading...
  2. 2026-11-12 Update README.md d34e468 4.5 KB
    Loading...
  3. 2026-11-12 Model card: tested API example; remove __pycache__ 91512e6 4.3 KB
    Loading...
  4. 2026-11-12 Add files using upload-large-folder tool fe8f8d6 3.9 KB
    Loading...
Catalog is the map. Apps are the tools.

Run models on your own machine, not in the cloud.

Every model page has an "Open in Abliteration" button that hands the model directly to the first-party desktop client, at the quantization your rig can actually run. No API keys, no subscription, no prompt leakage.

Open in Abliteration