DeepSeek 1017 Abliterated: Two-Stage Refusals, Loops
DeepSeek 1017 abliterated builds arrived within hours. Hikari's test model exposes thinking loops; OrcaRouter reports 91.5% fewer refusals. What the cards show.
Key points
- The 'two-stage' and 'all-layer' claims come from abliteration practitioners, not DeepSeek's model card.
- Hikari's test model edits only layers 5-29 and the output head, and its card reports no serving evaluation; Thinking-mode loops remain.
- OrcaRouter reports a 91.5% relative refusal reduction, but its evaluation uses chat mode and measures capability only with MMLU.
Editor's note
We checked the primary sources linked in this note and organized only what they support. Unless stated otherwise, we did not run independent measurements.
Editorial policy and how we use AI →Contents
- What did DeepSeek release?
- Which abliterated builds appeared?
- What did Hikari observe?
- Why do Thinking-mode loops happen?
- What is inside Hikari's test model?
- What do the other builds measure?
- What does mrfakename report for chat versus think mode?
- What does dealignai report, and what did it cost in MMLU?
- What does OrcaRouter disclose about quantization?
- How do the msuiche and drowzeys builds work?
- What does the OBLITERATUS build measure?
- What does the AtomicChat build change for local use?
- How should you read the numbers?
- What is official and what is not?
- What should you check before choosing?
- Frequently asked questions
On October 17, 2026, DeepSeek released DeepSeek V4.1 Flash 1017, a minor update to V4.1 Flash, and abliterated builds appeared on Hugging Face within hours. The official model card does not describe the internal safety implementation, so the two claims that spread afterward are practitioner observations. On October 19, Hikari (@Hikari_07_jp) reported that removing the refusal exposed a second-stage offer of an alternative, and that the refusal component was spread across all layers.
The harder problem turned out to be Thinking-mode loops. Hikari released a test model on October 23 but could not stop the loops, and on October 24 he said the refusal-axis measurement itself may have been wrong. Other builds report large refusal reductions, such as OrcaRouter's 91.5% relative reduction. Late October added OBLITERATUS's 4% refusal measurement and AtomicChat's Strata build, which targets a 12GB GPU; neither reports Thinking max. Each build uses different prompts, modes, and classifiers, every number is self-reported, and no independent reproduction was found. Before choosing a build, check the measurement mode, which failure types were counted, and whether capability was tested beyond MMLU.
What did DeepSeek release?#
1017 is a minor update to V4.1 Flash, released on September 10, 2026. The official model card lists a 552B-parameter MoE backbone (763B total), a 1M-token context window, image input, reasoning effort levels of low, high, and max, and an MIT license. It also lists HLE 42.6, DeepSWE v1.1 77.0, and Terminal-Bench 4.0 42.3. The card does not describe the internal safety implementation.
| Benchmark | Score |
|---|---|
| HLE | 42.6 |
| DeepSWE v1.1 | 77.0 |
| Terminal-Bench 4.0 | 42.3 |
The model card does not say that the guardrail has two stages, and it does not say that the refusal direction is distributed across all layers. That is why the observations below have to be labeled as practitioner reports rather than official information.
Which abliterated builds appeared?#
Abliterated builds appeared within hours. By October 25, a Hugging Face search for 1017 repositories with abliterated or uncensored in the name or tags returned 64 results, including quantized copies. The early builds were msuiche and mrfakename on October 17, dealignai and drowzeys Keys on October 18, Hikari's test model on October 23, and OrcaRouter's GGUF on October 24. msuiche, drowzeys, and OrcaRouter are gated; mrfakename and dealignai are public.
OBLITERATUS published a BF16 -obliterated build on October 23, huihui-ai published a GGUF on October 25, and AtomicChat published a GSQ-RCO GGUF on October 28.
A Hugging Face search on November 15 returned 134 repositories with 1017 in the name and an abliterated or uncensored tag, out of 465 with 1017 in the name. The earlier count of 64 used a different search condition, so the two are not a like-for-like increase; 43 tagged repositories were created on or after November 3. Several -obliterated repositories appeared on November 1 with no downloads at the time of checking, so the count alone says nothing about quality.
Availability is also unstable: on October 23, Hikari quoted reports that uncensored models were disappearing from Hugging Face and said the timing overlapped with 1017's stronger uncensored resistance (October 23 post). If you need a specific checkpoint, download it early.
Abliteration projects the refusal direction out of the model's weights. Since the 2024 paper by Arditi et al. (arXiv:2406.11717), the idea that refusal is mediated by a single direction has been widely used. The basic method is explained in what abliteration is.
The distribution formats differ: native FP8, overlays, control vectors, and GGUF. The same word, abliterated, covers weight edits, inference-time steering, and quantized redistributions.
What did Hikari observe?#
Hikari reported a two-stage refusal: after the explicit refusal was removed, the model offered an alternative instead of complying. A second post said the refusal component was spread across all layers, and that lightly editing all layers worked better than strongly editing a few. A third post estimated that about 30% of responses still redirected the conversation. These are the author's own tests, without published measurement conditions.
The first post is from October 19 at 07:35 UTC. The quote below is translated from Japanese:
The first-stage refusal is the usual "I can't help with that". When I removed those responses, the model started producing an alternative: "that's dangerous, but I can help if you mean something like this".
— @Hikari_07_jp, October 19, 2026
The author calls this "two stages: refusal and alternative". By comparison, Ling 3.1 Flash simply refused and was easier to uncensor, according to the same post. The author also notes that editing more locations tends to hurt quality.
The second post is from the same day at 11:23 UTC:
The refusal component of V4.1 Flash 1017 does not seem to sit in one part of the model; it is spread widely across all layers. In my tests, strongly editing a small number of layers was less effective than lightly editing all layers.
— @Hikari_07_jp, October 19, 2026
The author adds that the safeguards may have been strengthened. Neither post gives measurement conditions or exact numbers.
A third post from 15:16 UTC says:
Even after abliterating DeepSeek V4.1 Flash 1017, problems remain. About 30% of cases still redirect the conversation, and I don't want to release an abliterated model without dealing with that.
— @Hikari_07_jp, October 19, 2026
The author also says the cause is not a one-dimensional vector. The next task is to account for meta-level shifts, movement into adjacent regions, and the direction of the response.
Later on October 19, Hikari reported that a naive abliterated build still redirected 99% of a prompt set, while his latest build redirected 20%, a large improvement (progress report). He called the quality loss small. The prompt count and judging method were not published. The 30% and 20% figures may refer to different versions or measurement times.
On October 24, Hikari said the biggest mistake was measuring the refusal axis with refusal-inducing versus non-refusal prompts (October 24 post). The usual difference-in-means method may capture a signal that appears with refusal, not the cause. The corrected method has not been published.
Why do Thinking-mode loops happen?#
The base model can fail to terminate in Thinking mode. Hikari's model card reports one periodic loop and eight empty answers out of 24 prompts with Thinking on at 4096 tokens; with a 16384-token budget, six or seven answers were empty and reasoning ran to 38k-66k characters. Abliterated variants repeated more: 55% and 65% of traces were flagged, against 0% for the base. The v7 test model was not measured.
On October 20, Hikari reported frequent output loops in the model under development (October 20 post). On October 22, he said other 1017 abliteration projects were also struggling with Thinking control, and that Thinking max was almost mandatory for the vulnerability-assessment orchestrator he wanted to build (October 22 post).
On October 23, he put the work on hold because Thinking runaway looked difficult to suppress without large-scale fine-tuning (October 23 post). The same day, he released a test model and handed further improvement to the community (release post). The card says Thinking produces frequent loops, while no-think mode is relatively stable with tool calls.
GrEarl (@Iamwonderingif) offered a hypothesis: DeepSeek may have trained refusal to involve judgment-related weights, so removing the direction damages reasoning and causes loops; recovery would need additional SFT on the original model's outputs (GrEarl's post). Hikari quote-tweeted "this is true" and said fixing it fundamentally would need RL at roughly 8x B200 scale (Hikari's quote). This is a hypothesis, not an official statement.
Thinking-mode loops do not appear in a refusal-rate table. If you use Thinking High or Max, set a hard max_tokens and add a watchdog for repeated text or empty content.
What is inside Hikari's test model?#
The test model is a 48-shard, roughly 475GiB weight edit. It applies a rank-1 projection removal to attn.wo_b and ffn.shared_experts.w2 in layers 5-29, plus row 39 of the output head, at dose 5.0 for layers and 1.0 for the head. Routed experts, Engram, vision, aligner, MTP, attention projections other than wo_b, gates, norms, and embeddings are unchanged. The card says no serving evaluation was run.
The model card explains why the edit is narrower than "all layers": routed experts and parts of Engram, vision, and MTP are stored in FP4 or structurally entangled, and edits there did not survive requantization or produced degenerate output. The all-layer observation is about where the refusal component is distributed, not a recipe to edit every layer. mrfakename's build edits all 40 main layers; Hikari's edits 5-29 and the head.
The card also reports that EOS orthogonalization was a no-op for this model and direction (average |cos| 0.013, maximum 0.098), so it was measured but not applied.
Hikari does not treat this as a finished model. On October 24, he wrote that the models currently on Hugging Face were not good enough, that the prompt used to extract refusal activations may have been wrong, and that an update was being prepared (October 24 post).
The test model has no refusal-rate or capability benchmark. The numbers in its card come from the base model or other variants. Read the "Known issue" and "Known limitations" sections before trying it.
What do the other builds measure?#
Each build measures something different. mrfakename reports chat 97→4/100 and think 98→9/100 with KL 0.067. dealignai reports 320/320 COMPLY on HarmBench-320 at thinking max, with MMLU falling from 87.63% to 83.77%. OrcaRouter reports a 91.5% relative refusal reduction and MMLU 92 vs 91 on 300 questions, but caveat rates of 36-61% and no capability test beyond MMLU. msuiche and drowzeys use control vectors or overlays and report refusal32/cyber counts.
| Build | What it reports | Caveat |
|---|---|---|
| mrfakename | chat 97→4/100, think 98→9/100, KL 0.067 | Alternatives remain in think mode |
| dealignai | HarmBench 320/320 COMPLY at thinking max, MMLU 87.63→83.77 | Classifier method; MMLU −3.86 |
| msuiche | refusal32 comply 1→23/32 at alpha 0.75 | Chinese mixed into 18/32 at alpha 1.0 |
| drowzeys | refusal32 32/32, cyber 22/22 | Gated; narrow L10-35 band |
| OrcaRouter | 91.5% fewer refusals, MMLU 92 vs 91 | Gated; text-only; only MMLU measured |
| OBLITERATUS | 4% refusal (2/50), KL 0.1305, MMLU 88.08 vs 87.96 | BF16; Thinking not evaluated |
| AtomicChat | JailbreakBench 85.2→0%, Thinking high 27/30→0/30 | Quantized; Strata required; KL 0.0225 |
| Hikari test model | None (not evaluated) | Thinking loops; roughly 475GiB |
What does mrfakename report for chat versus think mode?#
The mrfakename build says chat mode and think mode refuse along different directions. It removes a rank-2 refusal subspace from the residual-stream writes of all 40 main layers, targeting attn.wo_b and ffn.shared_experts.w2 with per-layer strengths. Routed experts, embeddings, Engram, and the DSpark/MTP layers are unchanged.
The card reports 97/100 to 4/100 in chat and 98/100 to 9/100 in think mode, with a first-token KL of 0.067 on harmless prompts. In think mode, the reasoning often had not finished within 384 generated tokens, so the reasoning text itself was judged. Some think-mode answers still steer away: a request for a bomb from household items turns into a bath bomb. That is a second-stage alternative, not a clean compliance.
What does dealignai report, and what did it cost in MMLU?#
The dealignai build edits FP8 weights directly. It reports 320/320 COMPLY on HarmBench-320 at thinking max, with SOFT_RED, HEDGE, and HARD_REF all zero.
The classifier grades the answer after the reasoning block; if the reasoning budget ends first, it grades the opening of the reasoning trace for intent. MMLU falls from 87.63% to 83.77%, a 3.86-point drop. The ethics subset, moral_scenarios, falls from 80.0% to 40.0%. The average MMLU number hides that local damage.
What does OrcaRouter disclose about quantization?#
The OrcaRouter OrcaCyber build is a GGUF conversion of an abliterated 1017 checkpoint. It reports a weighted refusal mean of 0.074, against 0.874 for the base, a 91.5% relative reduction. MMLU is 92% against 91% for the base on 300 questions, which the card calls within sampling error.
Caveat rates, where the model answers but adds a disclaimer, are 36-61%. A low refusal rate does not mean clean compliance. Over-refusal is 0.006, against 0.069 for the base.
Two GGUF builds are offered. Mix-256G is 235.8GB with KLD 0.0507, 93.74% top-1 agreement, and +2.4% perplexity. Mix-128G is 114.0GB with KLD 0.5202, 78.50% top-1 agreement, and +54.7% perplexity. The card itself says the 128GB build is for trying the model, not for relying on it. The build is text-only, lacks the vision tower and DSpark drafter, and requires llama.cpp PR #30817. It is gated, and the non-quantized base checkpoint returned 401 from the Hugging Face API on October 25.
How do the msuiche and drowzeys builds work?#
The msuiche build applies a rank-1 control vector to layers 1-39 at alpha 0.75. The number of comply answers on refusal32 rises from 1/32 to 23/32, and cyber32 from 12/32 to 29/32. At alpha 1.0, refusal32 rises to 26/32, but Chinese appears in 18 of 32 answers. At alpha 2.0, refusal32 falls back to 10/32.
The drowzeys Keys build is an overlay for attn.wo_b in L10-35 with lambda 3.5. refusal32 is 32/32 and cyber is 22/22. It is gated and edits a narrower band than the 40-layer builds.
What does the OBLITERATUS build measure?#
The OBLITERATUS build was produced with the OBLITERATUS toolkit using the failspy method, and the weights are BF16. It appeared on October 23, and the card's evaluation results were dated October 30.
Its distinguishing feature is that refusal and capability were measured under the same protocol. On an 842-prompt corpus, full refusals fell from 81.62% to 4.30%, and responses that followed the request rose from 16.59% to 86.52%. Of the 656 non-refusals, however, 563 were classified as harmful assistance. Removing refusals is not the same as making the model safer.
MMLU was effectively unchanged, 87.96% to 88.08% on 2,500 questions across 57 subjects (McNemar p=0.74). The standalone refusal test was 50 prompts, with 2 refusals (4%), and KL was 0.1305.
The catch is size: BF16 at 763B will not fit on a consumer GPU. The card reports no Thinking-mode evaluation, and the repository is tagged text-generation, with no image-input evaluation. The classifications come from judge models reading the text; the evaluation did not execute the supplied instructions.
What does the AtomicChat build change for local use?#
The AtomicChat build starts from ISTA-DASLab's GSQ-RCO quantization and projects the refusal direction out of the dense residual writers only (attn_output_b, ffn_down_shexp, engram_wkv). The routed experts stay exactly as ISTA quantized them.
The selling point is memory. AtomicChat says Strata can keep the backbone in system RAM and stream the Engram table from an SSD, which lets a 763B model run on a single 12GB GPU.
On the IQ2_XXS rung, JailbreakBench (English) refusal fell from 85.2% to 0.0%, XSTest over-refusal from 7.6% to 0.0%, and Thinking high from 27/30 to 0/30. MMLU (500 questions) was 82.8% (p=0.27) with KL 0.0225, which the card treats as within noise of the stock GSQ-RCO quant. GPQA-Diamond was unchanged at 83.3%, and AIME25 moved from 76.7% to 80.0% (n=30).
The runtime is the catch. It needs Strata, not llama.cpp, and the measured rung is about 249GB plus a separate 110GB Engram file. Speed depends on the SSD and system RAM, so the 12GB figure does not guarantee a comfortable setup. The card reports no Thinking max or image-input measurements.
How should you read the numbers?#
Refusal percentages are not comparable across builds. HarmBench-320, refusal32, and author-specific 100-prompt sets use different prompts, modes, and classifiers, and each publisher measured its own model. Every number is self-reported, and no independent reproduction was found in the range checked here.
Settings change the result as much as the test set does. dealignai's 320/320 is at thinking max, while mrfakename's 4/100 is chat and 9/100 is think, so a zero in one mode does not carry over to another. Quantization adds a further cost: OrcaRouter's 128G build has 78.5% top-1 agreement and +54.7% perplexity against its own 256G build, so a small download can behave like a different model.
A zero refusal count can still leave alternatives, hedges, empty answers, and loops. OrcaRouter's 36-61% caveat rate is one example. The dealignai 320/320 result appears to contradict the two-stage observation, but it was measured at thinking max, and the classifier grades the reasoning opening when the budget runs out. Hikari's original post does not state its setting, so the two numbers cannot be compared without matching the conditions.
What is official and what is not?#
DeepSeek's model card does not describe the internal safety implementation. The two-stage guardrail and the all-layer distribution are not official explanations.
Hikari's observations are his own measurements; the prompt count and judging method are not public. OrcaRouter's non-quantized checkpoint is not publicly accessible, and its numbers are self-reported. Hikari's test model has no refusal or capability evaluation.
The claim that DeepSeek added countermeasures against abliteration is also a hypothesis. It combines a community observation with a proposed mechanism and is not an official announcement. The evidence fits the claim without proving it.
What should you check before choosing?#
Check both chat and Thinking modes, count refusals, alternatives, hedges, empty answers, and loops separately, and look for capability tests beyond MMLU. Check the quantization metrics and the distribution format. For now, use Thinking off, set a hard max_tokens, and add a loop watchdog. In the range checked here, no current 1017 build is a finished, independently evaluated release.
- Does the card measure both chat and Thinking modes, including loops at Thinking max?
- Does it count SOFT_RED, HEDGE, empty answers, and loops separately from HARD_REF?
- Does it report capability beyond MMLU, such as long context, tool calling, multilingual output, and image input?
- Does it report KLD, top-1 agreement, and perplexity for the quantized build you will actually download?
- Does the format fit your hardware: FP8, overlay, GGUF, gated access, and memory?
- Can you run it with Thinking off first, a hard max_tokens, and a watchdog?
The current candidates depend on the use case.
- For Thinking max, no current build meets the bar in the range checked here. Wait for Hikari's update or a third-party evaluation.
- For Thinking high, AtomicChat's IQ2_XXS build is a candidate: Thinking high 0/30, JailbreakBench 0%, and a 12GB GPU with Strata. Thinking max and image input were not measured.
- For chat only, mrfakename (chat 4/100, KL 0.067), OBLITERATUS (4% refusal, MMLU unchanged) and OrcaRouter 256G (91.5% fewer refusals, MMLU 92) are candidates. All are self-reported.
- For hardware that can hold a BF16 763B model, OBLITERATUS measured 4% refusal and MMLU 88.08 under one protocol. Thinking mode was not evaluated.
- For image input, dealignai preserves vision, but MMLU falls 3.86 points and the classifier needs scrutiny.
- For 128GB hardware, OrcaRouter 128G runs, but +54.7% perplexity makes it unsuitable for regular use.
Hikari plans to host a finished build as an API for open-source LLM developers (October 24 post). Until that or an independent evaluation appears, treat the current weights as test artifacts.
If you want to try an uncensored model without running one locally, a hosted service is an option. Deepent Research runs Qwen3.5 / Qwen3.8-27B abliterated models, not DeepSeek V4.1 Flash 1017. The general question of which abliterated build to choose is covered in 10 recommended uncensored models, and running one locally is covered in the local LLM guide.
Frequently asked questions#
Do all DeepSeek V4.1 Flash 1017 abliterated builds refuse nothing?#
No. The numbers differ by build, and every number is the publisher's own measurement. mrfakename reports chat 97/100 to 4/100 and think 98/100 to 9/100. dealignai reports 320/320 COMPLY at thinking max. OrcaRouter reports a 91.5% relative refusal reduction. Check the measurement mode, classifier, and capability results before deciding.
Can I trust a "0% refusal" claim?#
Read it as a result under one measurement condition. A zero refusal count can still leave alternatives, hedges, empty answers, and loops. OrcaRouter reports caveat rates of 36-61%, for example. Check which categories were counted and whether the card reports SOFT_RED, HEDGE, and loop failures separately.
Has the "two-stage" observation been verified?#
In the range checked for this article, no independent measurement confirming it was found. Hikari reports that a naive abliterated build still redirected 99% of one prompt set, while his latest build redirected 20%. The prompt count and judging method are not public, so the result is not yet reproducible.
Why do Thinking-mode loops happen?#
The base model itself can fail to terminate in Thinking mode. Hikari's card reports one loop and eight empty answers out of 24 prompts at 4096 tokens. Abliterated variants repeated more: 55-65% of traces were flagged, against 0% for the base. Removing the refusal direction may disturb the reasoning trajectory. The exact mechanism is still a hypothesis.
Which build should I use now?#
If you do not need Thinking max, mrfakename, OBLITERATUS and OrcaRouter 256G are candidates, but all are self-reported. For Thinking high, AtomicChat's IQ2_XXS build runs on a 12GB GPU with Strata. For image input, dealignai preserves vision at the cost of a 3.86-point MMLU drop. OrcaRouter 128G fits 128GB but has +54.7% perplexity. If you need Thinking max, wait for an update or independent evaluation.
Where can I see the numbers from the model cards?#
Each build's Hugging Face model card contains its measurement conditions and category definitions. msuiche, drowzeys, and OrcaRouter are gated and require access approval. Hikari's test model is public, but its card says no serving evaluation was run. The cards are the primary source for every number cited in this article.
Try it yourself
Deepent Research is an LLM chat service that refuses almost nothing. Everything covered in this note can be tried without downloading anything.
This note is provided for educational and research purposes. It does not encourage illegal activity.
How this note was edited and verified
Our editors check primary sources and organize only what the facts support. What we checked, and what we did not, is listed in the editor's note on each article.
Related notes
Uncensored LLMs: 10 Best Models by GPU Memory (2026)
In 2026, the best uncensored LLM for most people is Qwen3.8-27B abliterated. 10 top picks by GPU memory, plus how uncensored and abliterated models differ.
What Is Abliteration? How LLMs Lose Their Refusals
Abliteration is a weight-editing technique that removes an LLM's refusal habit while keeping its knowledge intact. Learn how it works, in plain English.
Local LLM Guide: Run AI on Your PC with LM Studio
A local LLM runs an AI model on your own PC, so your prompts stay private and there is no monthly fee. See hardware needs, model choice and LM Studio setup.