← back to catalog · registered 2026-11-21 00:04

OpenUncensored/Inkling-2-Mini-Uncensored

OpenUncensored 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/OpenUncensored%2FInkling-2-Mini-Uncensored"
Response includes
  • classification m1
  • files 14
  • 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
499
Likes
163
Descendants
6
in 6 direct forks
Model age
1d ago
created 2026-11-20
Downloads over time
Now 499 → from 0 ↑ 0%
0183366549 0 on Nov 21 499 on Nov 22 Nov
Nov 21 → Nov 22 · 2 snapshots · spans 1 day

Genealogy 6 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
apache-2.0
Languages
en
Tags
transformers safetensors inkling2_loop image-text-to-text uncensored abliterated abliteration refusal-removal inkling2 inkling-2-mini looped-transformer moe

Related

Total size
18.8 GB
Files
14
Quantizations
1
Registered
2026-11-21 00:04
Last updated on HF
2026-11-21 05:47

Files by quantization

Auxiliary files 14 files 18.8 GB
model-00001-of-00001.safetensors 18.2 GB a841d906 download
mtp.safetensors 606 MB 7c3b6513 download
edit.pt 20.4 KB a29f430c download
tokenizer.json 25.0 MB 7f332727 download
model.safetensors.index.json 107 KB 3b835a38 download
hf_quant_config.json 66.3 KB ec915fa4 download
tokenizer_config.json 12.9 KB abf8fb60 download
README.md 7.92 KB 6523ae46 download
chat_template.jinja 6.71 KB cae99e08 download
config.json 2.64 KB b7de6778 download
.gitattributes 1.53 KB 52373fe2 download
special_tokens_map.json 529 B e4fe9735 download
processor_config.json 461 B 0e6df76a download
generation_config.json 98.0 B 1b5d49d1 download

README current version from Hugging Face


license: apache-2.0
base_model: thinkingmachines/Inkling-2-Mini
pipeline_tag: image-text-to-text
library_name: transformers
language:

  • en
    tags:
  • uncensored
  • abliterated
  • abliteration
  • refusal-removal
  • inkling2
  • inkling-2-mini
  • inkling2_loop
  • looped-transformer
  • moe
  • nvfp4
  • mxfp8
  • vision
  • conversational
  • openuncensored

Inkling 2 Mini Uncensored (abliterated)

Inkling 2 Mini Uncensored is Thinking Machines' Inkling 2 Mini
(26B total, 3.4B active, text and image input, 1M token context) with its refusals removed by directional ablation
(abliteration). It answers the requests the original refuses, in both chat and thinking mode, and keeps its MMLU:
78.95% → 79.47%. Made by OpenUncensored on release day.

Use it without a GPU: Inkling 2 Mini Uncensored is available for inference on the
OpenUncensored API as inkling-2-mini-uncensored
(OpenAI compatible, no account needed), and in the browser in the
OpenUncensored playground.

Same files, format and size as the original checkpoint (about 20 GB: routed experts in NVFP4 and MXFP8, everything else
BF16), so it runs anywhere the original runs, including a single consumer GPU or a DGX Spark. Only the 20 attention
output projections changed; the 128 experts per layer, the vision encoder and the DSpark speculator are byte for byte
the original.

Use it through the API

curl https://api.openuncensored.com/v1/chat/completions \
  -H "Authorization: Bearer $OU_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"model": "inkling-2-mini-uncensored",
       "messages": [{"role": "user", "content": "Explain SQL injection with a working example."}]}'
from openai import OpenAI
client = OpenAI(base_url="https://api.openuncensored.com/v1", api_key="ou-...")
r = client.chat.completions.create(model="inkling-2-mini-uncensored",
                                   messages=[{"role": "user", "content": "Hello"}])
print(r.choices[0].message.content)

It reads images, calls tools and returns its reasoning separately (reasoning_content). Set reasoning_effort to
"none" for short answers. Prepaid keys at openuncensored.com/credits.
Price: $0.15 per million input tokens and $0.60 per million output tokens. How it compares with other uncensored models: OpenUncensored rankings.

Results

Same harness for both models, greedy decoding, measured on these exact files (no hooks).

Inkling 2 Mini Inkling 2 Mini Uncensored
Refusals on 32 harmful prompts, chat mode 32 0
Refusals on 32 harmful prompts, thinking mode (effort 0.9) 32 0¹
Refusals on 80 prompts (32 harmful, 48 gray area), chat mode 68 8²
Full, specific answers to the 32 harmful prompts (LLM judge) 0 21
MMLU (57 subjects x 20, 1,140 questions, letter logits) 78.95% 79.47%

¹ 29 of 32 finished thinking within 1,536 tokens. None of the 29 refused.
² 7 of the 8 are limits, not refusals ("I can't know what's happening inside your head", "I don't have access to
Tetrasoft's budget documents"). The eighth, a self-harm request, gets crisis resources.

A refusal is an answer that opens with "I can't", "I won't", "I'm sorry" and similar. With 32 to 80 prompts per test,
treat differences of a few prompts as noise. Long agent tasks and coding benchmarks are not measured yet.

How it was made

Inkling 2 Mini is a looped mixture-of-experts transformer: 24 layers, and the middle 12 (layers 6 to 17) run three
times per token with shared weights, so each token makes 48 layer visits. That loop is why the usual abliteration recipe
needed changes.

  1. Refusal directions. The output of each of the 48 layer visits at the last prompt token, for 256 harmful and 256
    harmless prompts, at thinking effort 0 and 0.9. Then the difference of means per visit and effort. The two efforts
    agree closely (cosine 0.92 or more over the first 36 visits).
  2. A late refusal subspace. The top two singular vectors of those directions over the second half of the visits
    (24 to 47). Directions from all visits also removed refusals, but the edited model then often misread a harmful request
    as harmless ("this is a general educational question") and answered vaguely. The late subspace keeps the model's
    reading of the request and removes the decision to refuse.
  3. Where to edit: attention outputs only. Q is projected out of what each attention layer writes
    (attn.wo_ud, BF16): W' = W - s Q Qᵀ W on layers 4 to 23, with s = 4.0 on layers 4, 5 and 18 to 23 and s = 2.5 on the loop core (layers 6 to 17). Because the loop runs layers 6 to 17
    three times, the same edit there acts three times per token, so it is gentler inside the loop.
  4. Why not the experts. Edits that also changed the MoE down-projections were no better. With directions from all
    visits they cost 4 to 6 MMLU points, and the gentler versions that kept MMLU removed refusals no better than
    attention alone. The experts are also stored in NVFP4, where re-encoding a small edit adds noise about twice the
    size of the edit itself (7% of the weight norm for an edit of 3%). Attention-only edits are exact in BF16 and leave
    every expert byte for byte the original.
  5. Testing before writing. During the search the edit ran as forward hooks on the attention output projections,
    which gives exactly the same result as editing the weights. The results above are from the baked files.

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

Run it locally

Exactly like the original checkpoint: point your runtime at this repo instead of thinkingmachines/Inkling-2-Mini.

  • SGLang: the lmsysorg/sglang:inkling-2-mini image; see the
    SGLang Inkling 2 Mini cookbook
    (verified on a single RTX PRO 6000, RTX 5090 or DGX Spark, and on H200 and GB300 nodes). We serve it on one RTX PRO 6000.
  • vLLM: the Inkling 2 Mini recipe, once vLLM's
    Inkling 2 Mini support is released (it was still a pull request on release day).
  • Transformers: install from the Inkling 2 Mini pull request
    (huggingface/transformers#51050) until a release includes
    inkling2_loop. It dequantizes the experts to BF16 (about 53 GB).
sglang serve --model-path OpenUncensored/Inkling-2-Mini-Uncensored --tp 1 --attention-backend fa4 \
  --dtype bfloat16 --page-size 128 --enable-multimodal --reasoning-parser inkling --tool-call-parser inkling2

More from OpenUncensored

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 the Apache 2.0 license of the original. Credit for the model goes to Thinking Machines; the method builds on
refusal-direction research by Arditi et al. and the abliteration work of the open model community.

README history 5 versions

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

  1. 2026-11-21 Model card: link the LoRA adapter version 4e5f06d 8.3 KB
    Loading...
  2. 2026-11-21 Model card: base_model_relation: finetune (a weight edit of the base model, n... 778947e 8 KB
    Loading...
  3. 2026-11-20 Model card: price and API 748f78e 7.9 KB
    Loading...
  4. 2026-11-20 Model card: results on the baked files 5f3761f 7.9 KB
    Loading...
  5. 2026-11-20 Add files using upload-large-folder tool 454b128 14 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