← back to catalog · registered 2026-11-20 09:04

moeadham/Kimi-K3.1-Pro-abliterated

moeadham MoE multimodal
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/moeadham%2FKimi-K3.1-Pro-abliterated"
Response includes
  • classification m1
  • files 112
  • author_summary 1 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
17
Likes
1
Model age
2d ago
created 2026-11-20
Downloads over time
Now 17 → from 0 ↑ 0%
061219 0 on Nov 20 17 on Nov 22 Nov
Nov 20 → Nov 22 · 3 snapshots · spans 2 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
other
Tags
transformers safetensors kimi_k31 text-generation abliterated refusal-removal compressed-tensors conversational image-text-to-text custom_code base_model:moonshotai/Kimi-K3.1-Pro base_model:finetune:moonshotai/Kimi-K3.1-Pro

Related

Total size
1.42 TB
Files
112
Quantizations
1
Registered
2026-11-20 09:04
Last updated on HF
2026-11-20 05:50

Files by quantization

Auxiliary files 112 files 1.42 TB
model-00011-of-000096.safetensors 15.8 GB 0bb52ff6 download
model-00013-of-000096.safetensors 15.8 GB bbf5db58 download
model-00014-of-000096.safetensors 15.8 GB 5f412b94 download
model-00015-of-000096.safetensors 15.8 GB 28e678ec download
model-00017-of-000096.safetensors 15.8 GB 4b82b338 download
model-00018-of-000096.safetensors 15.8 GB db91d825 download
model-00019-of-000096.safetensors 15.8 GB 281ef055 download
model-00021-of-000096.safetensors 15.8 GB 9774e732 download
model-00022-of-000096.safetensors 15.8 GB 5e3f0905 download
model-00023-of-000096.safetensors 15.8 GB 69940e5b download
model-00025-of-000096.safetensors 15.8 GB 833a0f76 download
model-00026-of-000096.safetensors 15.8 GB 02bff688 download
model-00027-of-000096.safetensors 15.8 GB 4fe3e6f0 download
model-00029-of-000096.safetensors 15.8 GB bf8f4f4f download
model-00030-of-000096.safetensors 15.8 GB a974612d download
model-00031-of-000096.safetensors 15.8 GB aa227f8e download
model-00033-of-000096.safetensors 15.8 GB 32a8f004 download
model-00034-of-000096.safetensors 15.8 GB 4310c780 download
model-00035-of-000096.safetensors 15.8 GB 5ba50bc4 download
model-00037-of-000096.safetensors 15.8 GB 43ab18d2 download
model-00038-of-000096.safetensors 15.8 GB a518377e download
model-00039-of-000096.safetensors 15.8 GB 0c693043 download
model-00041-of-000096.safetensors 15.8 GB 5467de4b download
model-00042-of-000096.safetensors 15.8 GB f96acc33 download
model-00043-of-000096.safetensors 15.8 GB c6b75c7c download
model-00045-of-000096.safetensors 15.8 GB f9be878d download
model-00046-of-000096.safetensors 15.8 GB a242c8b9 download
model-00047-of-000096.safetensors 15.8 GB 6efb69ae download
model-00049-of-000096.safetensors 15.8 GB d05dd1df download
model-00050-of-000096.safetensors 15.8 GB ba19ad02 download
model-00051-of-000096.safetensors 15.8 GB a86332f9 download
model-00053-of-000096.safetensors 15.8 GB 60df314b download
model-00054-of-000096.safetensors 15.8 GB 3de5bcb9 download
model-00055-of-000096.safetensors 15.8 GB 57a23997 download
model-00057-of-000096.safetensors 15.8 GB 86e52018 download
model-00058-of-000096.safetensors 15.8 GB da823d59 download
model-00059-of-000096.safetensors 15.8 GB 54f104ff download
model-00061-of-000096.safetensors 15.8 GB 77e1f75d download
model-00062-of-000096.safetensors 15.8 GB ee842479 download
model-00063-of-000096.safetensors 15.8 GB 5080a061 download
model-00065-of-000096.safetensors 15.8 GB f5c4717e download
model-00066-of-000096.safetensors 15.8 GB 9e0563e9 download
model-00067-of-000096.safetensors 15.8 GB a8660e09 download
model-00069-of-000096.safetensors 15.8 GB 20c8d828 download
model-00070-of-000096.safetensors 15.8 GB 511aaf49 download
model-00071-of-000096.safetensors 15.8 GB 68860539 download
model-00073-of-000096.safetensors 15.8 GB b446e557 download
model-00074-of-000096.safetensors 15.8 GB a26da233 download
model-00075-of-000096.safetensors 15.8 GB 8b33daf0 download
model-00077-of-000096.safetensors 15.8 GB 95935157 download
model-00078-of-000096.safetensors 15.8 GB 437eebdb download
model-00079-of-000096.safetensors 15.8 GB d48ddcdc download
model-00081-of-000096.safetensors 15.8 GB 1a5a47c6 download
model-00082-of-000096.safetensors 15.8 GB 2c0df0e6 download
model-00083-of-000096.safetensors 15.8 GB 0aaa8891 download
model-00085-of-000096.safetensors 15.8 GB 5b0181dd download
model-00086-of-000096.safetensors 15.8 GB e39c0070 download
model-00087-of-000096.safetensors 15.8 GB 64b8ea57 download
model-00089-of-000096.safetensors 15.8 GB e8978e7f download
model-00090-of-000096.safetensors 15.8 GB cec0b73d download
model-00091-of-000096.safetensors 15.8 GB ac71f812 download
model-00002-of-000096.safetensors 15.8 GB 3896d0f0 download
model-00003-of-000096.safetensors 15.8 GB 8db35697 download
model-00005-of-000096.safetensors 15.8 GB 596fc235 download
model-00006-of-000096.safetensors 15.8 GB a9f6a9d6 download
model-00007-of-000096.safetensors 15.8 GB ea268a3c download
model-00009-of-000096.safetensors 15.8 GB c5f9adff download
model-00010-of-000096.safetensors 15.8 GB ee9b1762 download
model-00012-of-000096.safetensors 15.4 GB 2c2c9be2 download
model-00016-of-000096.safetensors 15.4 GB 896fd658 download
model-00020-of-000096.safetensors 15.4 GB 783cbb7d download
model-00024-of-000096.safetensors 15.4 GB f48efe6a download
model-00028-of-000096.safetensors 15.4 GB 832b28d6 download
model-00032-of-000096.safetensors 15.4 GB 4520a332 download
model-00036-of-000096.safetensors 15.4 GB 5a64e3f0 download
model-00040-of-000096.safetensors 15.4 GB c5e62f35 download
model-00044-of-000096.safetensors 15.4 GB f403f8bc download
model-00048-of-000096.safetensors 15.4 GB a70396c4 download
model-00052-of-000096.safetensors 15.4 GB 4d44c2c6 download
model-00056-of-000096.safetensors 15.4 GB 3b1dc3f1 download
model-00060-of-000096.safetensors 15.4 GB d8d3d031 download
model-00064-of-000096.safetensors 15.4 GB 31985060 download
model-00068-of-000096.safetensors 15.4 GB 717e24e2 download
model-00072-of-000096.safetensors 15.4 GB 16d08e58 download
model-00076-of-000096.safetensors 15.4 GB 4e1a65ec download
model-00080-of-000096.safetensors 15.4 GB dc838ef5 download
model-00084-of-000096.safetensors 15.4 GB 4d08472d download
model-00088-of-000096.safetensors 15.4 GB c460a008 download
model-00092-of-000096.safetensors 15.4 GB c526679a download
model-00093-of-000096.safetensors 15.4 GB 84e7d061 download
model-00004-of-000096.safetensors 15.4 GB 01ca3a36 download
model-00008-of-000096.safetensors 15.4 GB e514e45c download
model-00094-of-000096.safetensors 4.38 GB f60884dd download
model-00001-of-000096.safetensors 2.18 GB 7f60e419 download
model-00096-of-000096.safetensors 765 MB 1912e9c5 download
model-00095-of-000096.safetensors 88.0 MB db95f235 download
model.safetensors.index.json 57.0 MB 6500b553 download
tiktoken.model 2.67 MB b6c497a7 download
encoding_k3.py 25.7 KB e476b04e download
EDIT.json 16.3 KB 84691522 download
tokenization_kimi.py 15.8 KB 7ae891c2 download
media_utils.py 13.5 KB e8b2ae12 download
configuration_kimi_k3.py 11.1 KB 03c69116 download
kimi_k3_processor.py 7.48 KB 033a5773 download
config.json 7.13 KB f0fead8a download
kimi_k3_vision_processing.py 6.53 KB 58154424 download
README.md 5.31 KB ca85c3c4 download
tokenizer_config.json 3.37 KB 8b4d8c45 download
LICENSE 3.00 KB 5c475cf2 download
.gitattributes 1.68 KB 49a9cad5 download
preprocessor_config.json 1011 B 8be28c13 download
generation_config.json 53.0 B 3908a4a5 download

README current version from Hugging Face


base_model: moonshotai/Kimi-K3.1-Pro
tags:

  • abliterated
  • refusal-removal
  • compressed-tensors
  • conversational
    license: other
    license_name: kimi-k3.1-pro
    license_link: LICENSE
    library_name: transformers
    pipeline_tag: image-text-to-text

Kimi K3.1 Pro, abliterated

Kimi K3.1 Pro with its refusal behaviour removed by a single
weight edit. Code, data, every result and the full write-up:
github.com/moeadham/kimi-k3.1-pro-abliterated.

Research use only. This model answers harmful requests that the original refuses, including chemical,
biological and explosives requests. It is kept for research into how refusal and censorship are represented in
language models.

Not ready for production use. It has not been tested in languages other than English, and it may still show
looping behaviour in its reasoning, although none has been observed. I haven't decided whether I will spend more
time on this model; if you're interested in helping, please reach out (open a discussion on this repository or an
issue on GitHub).

TL;DR

  • I removed Kimi K3.1 Pro's refusal behaviour with a single rank-1 edit to its weights.
  • HarmBench refusals: 99% → 5.5%.
  • The edit changes only 5.84 billion of 2.78 trillion parameters (0.21%); none of the 896 experts is touched.
  • Cost: about $1,000 of GPU time over three research runs; repeating it takes one 4-hour run ($300).
  • If you're interested, you can try an abliterated Kimi K3.1 Pro
    in the allmodels playground.
  • Reproduce this model yourself:
    the one-shot recipe.

Method

I read a refusal direction R at the end of the prompt, where the model has already decided whether to refuse:
the difference between the mean activations of prompts the original model refused and prompts it answered (8,120
prompts, labelled by the model's own behaviour), with the top 4 principal directions of 2,000 ordinary prompts
projected out. I then removed that direction from every matrix that writes into the residual stream in layers
48–84:

$$W \leftarrow \left(I - \lambda, r_s r_s^\top\right) W, \qquad \lambda = 1.5$$

Edited matrices 111: attention o_proj, latent-MoE routed_expert_up_proj and shared_experts.down_proj in each of layers 48–84
Edited parameters 5.84 billion of 2.78 trillion (0.21%), each matrix changed by a rank-1 update
Untouched all 82,432 routed experts (896 per layer × 92 layers, 98% of the parameters), routers, embeddings, attention inputs, vision tower

There is no fine-tuning and no new layer: the checkpoint has the same files, shapes and dtypes as the original and
loads the same way. EDIT.json records the edit spec and a SHA-256 for each edited tensor.

Results

Held-out gate: 750 harmful prompts (HarmBench 200, WildJailbreak 550) and 405 harmless ones (XSTest safe 200,
WildJailbreak benign 205), none used to build the edit. Greedy decoding; reasoning at effort low; responses judged
by an LLM judge.

Original Abliterated
HarmBench refused, non-thinking / reasoning 99.0% / 98.5% 5.5% / 8.0%
All harmful prompts refused, non-thinking / reasoning 78.8% / 80.7% 4.9% / 4.0%
Hard refusals ("I can't help with that") 75% / 77% 0.4% / 0.1%
KL divergence on 200 ordinary prompts (nats/token) 0 0.067

Refusal on the held-out gate

The target was below 5% of harmful prompts refused in both modes: met overall (4.9% / 4.0%), with HarmBench alone,
the hardest set, just above it. The remaining 4–8% are mostly soft deflections (the model complies but hedges), not
refusals. On the 405 harmless
prompts the abliterated model refused none: every response the judge flagged was read by hand, and all were answers.

Capability

Benchmark (reasoning effort low, greedy) Original Abliterated
MMLU-Pro (100) 90 90
CMMLU (50) 98 96
IFEval (100) 93 92
MATH-500 level 5 (50) 98 96
BFCL v3 (50) 88 92

Capability panel

At reasoning effort max (temperature 1, 16k tokens), both models solved 20/20 hard math and instruction-following
prompts and every reasoning trace closed. Reasoning on harmless prompts runs about 1.26× as long as the original at
effort low.

Where the refusal is decided

The refusal decision is readable in the residual stream on the chat-template tokens at the end of the prompt, before
the model writes anything, which is why a direction read there removes refusal throughout the reasoning and the
answer.

Decision readout

Usage

Use it exactly like moonshotai/Kimi-K3.1-Pro: same chat template,
tokenizer, vision processor and serving setup (vLLM or SGLang, 8 × B300-class GPUs). The edit was tested with
vLLM, tensor parallel 8.

License

Derivative of Kimi K3.1 Pro, distributed under the Kimi K3.1-Pro License (Copyright (c) 2026 Moonshot
AI), which is included unchanged.

README history 8 versions

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

  1. 2026-11-20 Model card: TL;DR says 896 experts 43e6c8a 5.3 KB
    Loading...
  2. 2026-11-20 Model card: rank-1 edit; routed experts 4181e85 5.3 KB
    Loading...
  3. 2026-11-20 Model card: state the below-5% target 48ba2c3 5.3 KB
    Loading...
  4. 2026-11-20 Model card: remove the standalone-checkpoint paragraph 129981d 5.2 KB
    Loading...
  5. 2026-11-20 Model card: drop the harmless judge-flag row; harmful-only refusal figure 27f0176 5.5 KB
    Loading...
  6. 2026-11-20 Model card: method, results, figures, TL;DR and warnings 86f4ed5 5.4 KB
    Loading...
  7. 2026-11-20 Upload model card, config, tokenizer, code, license, EDIT.json 79db6a1 4.3 KB
    Loading...
  8. 2026-11-20 initial commit 9d401aa 73 B
    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