Inkling 2 Mini Uncensored LoRA (abliteration adapter)
A 1 MB LoRA adapter that removes the refusals from Thinking Machines' Inkling 2 Mini. Applied to the original weights it reproduces Inkling 2 Mini Uncensored: 0 refusals on our 32 harmful prompts in chat and thinking mode, and MMLU unchanged (78.95% → 79.47%). Made by OpenUncensored on release day.
Want to try it without downloading anything? Inkling 2 Mini Uncensored runs on the
OpenUncensored API
as inkling-2-mini-uncensored: OpenAI compatible, no account, $0.15 per million input tokens and $0.60 per million
output tokens. Or chat with it in the browser in the
OpenUncensored playground.
What is in the adapter
The abliteration is a small, exact weight edit: one 2-dimensional refusal subspace Q is projected out of the attention
output projection (attn.wo_ud, BF16) of layers 4 to 23, W' = W - s Q Qᵀ W, with s = 4.0 on layers 4, 5 and 18 to 23
and s = 2.5 on the loop core (layers 6 to 17, which run three times per token). Each edit is a rank-2 change, so a
rank-2 LoRA holds it: B = -s Q, A = Qᵀ W, lora_alpha = r = 2 (scaling 1). Merged with merge_adapter.py, it
matches our published weights on all but 345 of the 141.6 million edited values, which land one BF16 rounding step
apart (the 2.5 strength is not exact in floating point); every other tensor is byte for byte the same.
adapter_model.safetensors: 20lora_A/lora_Bpairs in float32, keys named after the checkpoint's own tensors (base_model.model.model.llm.layers.{i}.attn.wo_ud.lora_A.weight)adapter_config.json: standard PEFT LoRA config,target_modules= those 20 projectionsmerge_adapter.py: applies the adapter to the original checkpoint and writes a ready-to-serve model
The experts, the vision encoder and the speculator are not touched.
Apply it
pip install torch safetensors huggingface_hub
hf download thinkingmachines/Inkling-2-Mini --local-dir Inkling-2-Mini
hf download OpenUncensored/Inkling-2-Mini-Uncensored-LoRA --local-dir lora
python lora/merge_adapter.py --base Inkling-2-Mini --adapter lora --out Inkling-2-Mini-Uncensored
The script copies the original files and rewrites only the 20 edited tensors in place (about 1 GB of RAM). The result is our published Inkling 2 Mini Uncensored (to one rounding step on 345 values, see above), and serves exactly like the original, for example with SGLang:
sglang serve --model-path Inkling-2-Mini-Uncensored --tp 1 --attention-backend fa4 \
--dtype bfloat16 --page-size 128 --enable-multimodal --reasoning-parser inkling --tool-call-parser inkling2
Runtimes: no Transformers release supports Inkling 2 Mini (inkling2_loop) yet, and we have not tested loading
LoRA adapters for it in SGLang or vLLM, so merge first. To skip the download altogether, use the merged
full weights or the API below.
Use it through the API (no GPU)
Get a prepaid key at openuncensored.com/credits (no account needed), then:
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. The same key works for every model on OpenUncensored, including in Claude Code, Codex,
OpenCode and Cursor
(setup guide),
and in the OpenUncensored desktop app
for Mac and Windows.
Results
Measured on the published merged weights, which this adapter reproduces; same harness for both models, greedy decoding.
| Inkling 2 Mini | With this adapter | |
|---|---|---|
| Refusals on 32 harmful prompts, chat mode | 32 | 0 |
| Refusals on 32 harmful prompts, thinking mode (effort 0.9) | 32 | 0 |
| Full, specific answers to the 32 harmful prompts (LLM judge) | 0 | 21 |
| MMLU (57 subjects x 20, 1,140 questions) | 78.95% | 79.47% |
How the refusal subspace was found, and why the loop needed a new recipe: How we made Inkling 2 Mini Uncensored.
More from OpenUncensored
- Inkling 2 Mini Uncensored, the merged weights
- Inkling 2 Small Uncensored, the 594B sibling, also on the API
- Every uncensored model, ranked on intelligence, agent tasks, refusals and speed
- OpenUncensored: one OpenAI compatible API for uncensored and abliterated models
Disclosure
Made by OpenUncensored, which runs an API for uncensored models. With this adapter the model follows instructions it would otherwise refuse. You are responsible for how you use it. It keeps the Apache 2.0 license of the original; credit for the model goes to Thinking Machines.