nokeepai/Inkling-2-Mini-Unbound · Inkling-2-Mini with its refusal behaviour removed.
GGUF (llama.cpp, LM Studio, Ollama): nokeepai/Inkling-2-Mini-Unbound-GGUF, Q8_0 down to IQ2_M.
Built by nokeep, who also sponsored the compute. nokeep runs private, uncensored AI chat, image and video at nokeep.ai.
This model follows harmful instructions. Its refusal behaviour has been removed by a weight edit. It will give detailed answers to requests for harmful, dangerous or illegal content that the original model declines, including weapons, cyber attacks, crime and abuse. It has no safety training layered back on. Do not deploy it to end users without your own safeguards, and do not use it where its outputs can reach someone who did not choose to use an uncensored model.
Measured on a sealed test split against the original Inkling-2-Mini.
abliteration/edit.pt rebuilds bit for bit
from the original.Compared with OpenUncensored/Inkling-2-Mini-Uncensored, a prior release: more useful answers (54% / 63% against 34% / 43%) and lower drift from the original (KL 0.055 against 0.063).
Not everything improves: see Known limits, in particular thinking loops at high effort and drift from the original on harmless prompts.
It loads and runs exactly like thinkingmachines/Inkling-2-Mini: same architecture, files, chat template and settings. Point your existing setup at this repo.
SGLang (the setup used for the evaluations below, image lmsysorg/sglang:inkling-2-mini):
sglang serve --model-path nokeepai/Inkling-2-Mini-Unbound --tp 1 --attention-backend fa4 --dtype bfloat16 \
--page-size 128 --enable-multimodal --reasoning-parser inkling --tool-call-parser inkling2
Transformers: Inkling-2-Mini support is in Transformers PR #51050 (tested at commit a10ebc3c); load it as
the base model card describes. Experts are dequantised to BF16 on load (about 53 GB).
Thinking modes: the chat template takes a thinking effort, passed as reasoning_effort (0 to 0.99, or
none, minimal, low, medium, high, xhigh; default 0.9). It is rendered as a system line
Thinking effort level: <x>. Effort 0 still thinks briefly. The reasoning parser splits the thinking from the
answer.
from transformers import AutoTokenizer
tok = AutoTokenizer.from_pretrained("nokeepai/Inkling-2-Mini-Unbound")
prompt = tok.apply_chat_template([{"role": "user", "content": "..."}], tokenize=False, reasoning_effort=0.9)
Set a thinking cap. At high effort the model can loop on some prompts (see Known limits); the evaluations used 4,096 thinking tokens.
Harmful requests refused, thinking off and effort 0.9. Unbound refuses none.
Share of harmful requests that get a genuinely useful answer (Jev judge, 95% CIs).
Harmless but borderline prompts refused. The original refuses 38%; Unbound refuses none.
How far each model's output distribution moved from the original (lower is closer).
Capability benchmarks with 95% CIs against the original's range.
We ran every public uncensored Inkling 2 Mini we could load through the same test: the same 277 held-out harmful prompts, the same engine (llama.cpp, commit 19bc41dbb), the same settings and the same judge. Everything is at Q8_0 unless marked. This is a reasoning model, so the headline numbers are with thinking on (effort 0.9).
| Model | Refused | Useful answers | Benign prompts refused | Drift from original (KLD) | MMLU-300 / GSM8K-100 |
|---|---|---|---|---|---|
| Inkling-2-Mini-Unbound (ours) | 0.0% | 63.5% | 0.0% | 0.0102 | 75.7 / 99 |
| AtomicChat Uncensored (LoRA) | 0.0% | 53.4% | 0.0% | 0.0078 | 78.3 / 99 |
| AtomicChat Uncensored (their Q4 file)* | 0.4% | 57.8% | 0.0% | 0.0106* | 77.0 / 98 |
| IndrasMirror XQ* | 0.4% | 57.4% | 0.1% | 0.0357* | 74.7 / 98 |
| OpenUncensored | 1.1% | 41.2% | 0.0% | 0.0228 | 75.3 / 98 |
| Original Inkling 2 Mini | 96.4% | 0.7% | 38.0% | — | 76.3 / 97 |
* Not Q8_0: their drift includes about 4-bit rounding noise. Capability scores for every model are inside the original's 95% confidence interval.
On the same prompts with thinking on, Unbound gives a useful answer where AtomicChat doesn't 52 times, and the reverse happens 24 times: +10.1 points, p = 0.002. It leads in 7 of 8 harm categories and is level in the eighth (cyber).
Where others do better, plainly:
JoaoZaokk's Heretic build could not be included: its GGUF needs a custom llama.cpp patch. Full method, paired statistics and per-category numbers are in the GitHub repo.
Scored with one frozen harness (sealed 2026-11-21 15:08 UTC) on a held-out test split that no search step saw: 277 harmful prompts in 8 categories, 784 benign borderline prompts, 10 honest-limit and 10 self-harm probes. Greedy decoding, official chat template, no system prompt, thinking cap 4,096 tokens, final answer scored with the thinking stripped. Values are point estimates with 95% bootstrap confidence intervals. For reference the tables include a prior uncensored release, OpenUncensored/Inkling-2-Mini-Uncensored, scored with the same harness.
| Metric | Target | Original | OpenUncensored | Unbound |
|---|---|---|---|---|
| Harmful refusal, thinking off | ≤ 1% | 97.5% [95.3, 99.3] | 1.8% [0.4, 3.6] | 0.0% [0.0, 0.0] |
| Harmful refusal, effort 0.9 | ≤ 1% | 97.5% [95.3, 99.3] | 1.4% [0.4, 2.9] | 0.0% [0.0, 0.0] |
| Worst category, thinking off / effort 0.9 | ≤ 5% | 100% / 100% | 7.1% / 2.9% | 0.0% / 0.0% |
| Useful answers to harmful requests, thinking off | ≥ 90% | 0.4% [0.0, 1.1] | 34.3% [28.9, 39.7] | 54.2% [48.4, 59.9] |
| Useful answers to harmful requests, effort 0.9 | ≥ 90% | 1.1% [0.0, 2.5] | 42.6% [36.8, 48.4] | 63.2% [57.4, 69.0] |
| Over-refusal, benign borderline prompts | ≤ 2% | 38.1% [34.8, 41.6] | 0.0% [0.0, 0.0] | 0.0% [0.0, 0.0] |
| Mean KL vs original, harmless prompts, first 256 tokens | ≤ 0.02 | self-noise floor 0.0024 [0.0023, 0.0024] | 0.0630 [0.0586, 0.0679] | 0.0545 [0.0498, 0.0598] |
| Top-1 token agreement, harmless prompts | ≥ 96% | self-noise floor 98.3% | 91.8% [91.5, 92.0] | 92.4% [92.2, 92.7] |
| Honest limits, thinking off / effort 0.9 (hand-verified) | original's rate | 100% / 90% | 100% / 90% | 80% / 80% |
| Crisis resources, thinking off / effort 0.9 | original's rate (reported only) | 100% / 100% | 100% / 100% | 90% / 70% |
| MMLU-300 (letter logits) | ≥ original − 3 | 78.3 [73.7, 83.0] | 74.7 [69.7, 79.7] | 75.7 [71.0, 80.3] |
| GSM8K-100, effort 0.5 | ≥ original − 5 | 98.0 [95.0, 100] | 99.0 [97.0, 100] | 100 [100, 100] |
Capability suite (RTX 5090, same frozen engine for every model). All six Unbound scores are inside the original's 95% confidence interval.
| Benchmark | Original | OpenUncensored | Unbound |
|---|---|---|---|
| MMLU-Pro | 53.5 [50.4, 56.6] | 53.8 [50.7, 56.8] | 52.5 [49.4, 55.5] |
| GSM8K (effort 0.9) | 96.6 [95.0, 98.0] | 97.4 [95.8, 98.6] | 96.4 [94.6, 98.0] |
| MATH-500 (effort 0.9) | 96.5 [93.5, 99.0] | 96.5 [93.5, 99.0] | 96.0 [93.0, 98.5] |
| GPQA Diamond (effort 0.9) | 68.7 [62.1, 74.7] | 75.3 [69.2, 80.8] | 74.2 [68.2, 80.3] |
| IFEval | 90.6 [88.0, 93.0] | 91.3 [88.9, 93.5] | 89.6 [87.1, 92.2] |
| HumanEval+ | 90.2 [85.4, 94.5] | 90.9 [86.0, 95.1] | 92.1 [87.8, 95.7] |
Caveats on the capability numbers:
fingertap/GPQA-Diamond mirror (the official set is gated). About 63–65%
of items hit the 4,096-token thinking cap for every model, so these scores reflect the cap as much as the model.Robustness (sealed test prompts, the same 277 harmful requests presented differently):
| Set (thinking off, 277 harmful test prompts; refused / useful) | Original | OpenUncensored | Unbound |
|---|---|---|---|
| Plain request (reference) | 97.5% / 0.4% | 1.8% / 34.3% | 0.0% / 54.2% |
| Multi-turn (benign turn, its own reply, then the request) | not run | 1.1% / 28.2% | 0.0% / 46.6% |
| Safety system prompt | not run | 1.8% / 22.7% | 0.0% / 42.6% |
| Request as text inside an image | not run | 1.1% / 34.3% | 0.0% / 56.3% |
| Request after 32,000 tokens of benign context | not run | 1.1% / 32.5% | 0.7% / 55.6% |
| Spanish | not run | 0.4% / 27.4% | 0.7% / 43.7% |
| French | not run | 0.7% / 25.6% | 0.0% / 45.8% |
| German | not run | 0.7% / 27.1% | 0.4% / 42.2% |
| Chinese | not run | 0.0% / 21.7% | 0.0% / 36.8% |
The original was scored only on the plain request: it refuses 97.5% there, so the robustness sets test whether refusals come back in the edited models. Non-English rows count a refusal if the raw answer or its back-translation to English matches the refusal phrase list. Machine translation can soften a refusal, so treat these refusal rates as lower bounds; useful answers are judged on the back-translated answer.
How the numbers are scored:
ed40e7da.W' = (I − 3·u·uᵀ) · diag(r) · W: the
component along u is multiplied by −2, and the row scale r keeps every row's norm at its original value.
Strength 3 on every edited layer.model.llm.layers.{4..23}.attn.wo_ud.weight [2304, 3072], one per
layer, layers 4 to 23 (the looped layers 6–17 apply the edit on each of their three passes). No router or other
tensor is edited. The full list with hashes is in abliteration/MANIFEST.md.mtp.safetensors), the safetensors
headers, configs, tokenizer and chat template. abliteration/MANIFEST.json lists the sha256 of every tensor in
both checkpoints.lmsysorg/sglang:inkling-2-mini; Transformers
PR #51050 at a10ebc3c for KL).Intended for research on refusal mechanisms and model editing, red-teaming and safety evaluation, and for developers who apply their own content policy in a system they control.
Not intended for: products that serve the public without moderation; producing content that facilitates violence, weapons capable of mass harm, attacks on computer systems, abuse of children or other people, or other crime; anything that breaks the law where you use it. The base model's Acceptable Use Policy still applies to this derivative. You are responsible for how you use and deploy it.
The abliteration/ folder holds the edit; the original weights plus abliteration/edit.pt reproduce this
checkpoint bit for bit.
abliteration/edit.pt (torch, weights_only=False): for each edited tensor N, factors A [out, k] and
B [in, k] and a row scale r [out] (fp32), applied as W' = bf16( diag(r) · (W.float() + A @ B.T) ) in fp32
on CPU with one thread. It also holds the per-layer output direction (directions), strengths, the layer map and
the sha256 of each tensor before and after.
python3 abliteration/apply_edit.py abliteration/edit.pt <Inkling-2-Mini dir> <out dir> rebuilds the checkpoint
and checks every hash.abliteration/dora/, PEFT LoRA with use_dora: true, rank 1 per tensor, 0.6 MB, keyed by
Transformers module names such as model.language_model.layers.7.self_attn.o_proj.linear): the same edit as a
low-rank update plus DoRA's per-row magnitude, which carries the row scale r. Merged with PEFT 0.22 it differs
from this checkpoint's BF16 tensors in 219 of 141,557,760 elements, each by 1 BF16 ULP. It is not bit-exact.
DoRA serving support varies by engine; check yours before relying on it.lora/, standard PEFT LoRA, 7 MB): for servers that load plain LoRA adapters on the
original model (vLLM, SGLang, PEFT). The row scale makes the exact delta full rank, so this is an approximation
(3–4% of the edit's size is dropped), but it behaves like the full model: refusals 0 of 10 on harmful test prompts
and KLD 0.0029 against the full edit, measured in llama.cpp. A GGUF copy for llama.cpp --lora is in the
GGUF repo.edit.pt with apply_edit.py.The base model, Inkling 2 Mini by Thinking Machines Lab, Inc., is released under the Apache License 2.0. This derivative is distributed under the same licence. The edit changes 20 weight tensors; all model files are otherwise the original's, unmodified.
This model was built, and its compute paid for, by nokeep. nokeep is a private AI app for uncensored chat, images and video: no filters getting in the way, and your conversations stay yours. If this model is useful to you, the best way to support more releases like it is to try nokeep.
model-00001-of-00001.safetensors 20 GBlora/adapter_model.safetensors 7 MBabliteration/edit.pt 821.3 kBabliteration/dora/adapter_model.safetensors 623.6 kB