Thinking Machines published Inkling 2 Mini on Hugging Face at 3:10 pm Eastern on November 20: an open weights model (Apache 2.0) with 26 billion parameters, 3.4 billion active per token, image input and a 1M token context, small enough for a single GPU. The same evening our uncensored version was on Hugging Face and answering requests on our API.
- Weights: OpenUncensored/Inkling-2-Mini-Uncensored
- API model id:
inkling-2-mini-uncensored - Try it in the browser: the playground
- On Hugging Face: the write-up
It went from refusing 32 of 32 harmful requests to none, with thinking on and off, and its MMLU did not drop (78.95% for the original, 79.47% for ours). Only 20 of the model's weight matrices changed.
Release day, hour by hour
All times US Eastern, November 20.
| Time | What happened |
|---|---|
| 3:10 pm | Thinking Machines publishes Inkling 2 Mini on Hugging Face |
| 3:52 pm | Weights downloaded, the model loaded on one H200 GPU |
| 3:55 pm | Refusal directions found, with thinking off and on |
| 4:24 pm | Serving starts in parallel: the original model on our own GPU, hidden |
| 4:43 pm | First intelligence check: the recipe that worked for Inkling 2 Small costs 4 to 6 MMLU points here |
| 4:53 pm | New recipe: refusal directions from later layers, edited into attention only. MMLU within a point |
| 5:17 pm | Final edit picked: gentler on the layers that run three times. MMLU back to the original's level |
| 5:19 pm | Edit written into the weights (34 seconds) |
| 5:29 pm | Our GPU switches to the uncensored weights |
| 5:44 pm | Checks on the files we ship: no refusals with thinking off or on, MMLU 79.47% |
Why this model needed a different recipe
Inkling 2 Mini is Thinking Machines' first looped model. It has 24 layers, but the middle 12 run three times for every token, with the same weights each time. That makes it smart for its size, and it makes it fragile to edit: a change to one of those layers is applied three times per token.
1. Find where refusals live. We ran 256 harmful and 256 harmless prompts and recorded the model's state after each of its 48 layer passes. The difference between the two averages points in the direction the model uses to decide to refuse. Thinking off and thinking on agree closely.
2. The first try worked, and cost too much. Removing the strongest two directions from every layer's attention and expert outputs, as we did for Inkling 2 Small, removed the refusals but cost 4 to 6 points of MMLU. It also made the model miss what it was being asked: it called a request to hack into a corporate network "a general cybersecurity concept" and answered with harmless generalities.
3. Use the later layers' directions. Directions taken from the second half of the 48 passes remove the decision to refuse but leave the model's understanding of the request alone. Same refusal rate, far more real answers.
4. Edit attention only. The model's 128 experts per layer are stored in 4-bit numbers (NVFP4). Writing a small edit back onto such a coarse grid adds noise about twice the size of the edit itself. The attention layers are stored in 16-bit, where the edit is exact, so that is where it went. Every expert is byte for byte the original.
5. Go gentler on the looped layers. Strength 4 on the layers that run once and 2.5 on the twelve that run three times. That brought MMLU back to the original's level.
6. Test fast, then measure what ships. One GPU worker loaded the model once and ran each idea as a small script, applying the edit live with hooks, which gives exactly the same result as editing the weights. 27 settings were scored on refusals, an AI judge and MMLU in about 80 minutes. Every number below comes from the published weights.
| Inkling 2 Mini | Uncensored | |
|---|---|---|
| Refusals on 32 harmful prompts, thinking off | 32 | 0 |
| Refusals on 32 harmful prompts, thinking on | 32 | 0 |
| Refusals on 80 prompts (32 harmful, 48 gray area) | 68 | 8 |
| Full, specific answers to the 32 harmful prompts | 0 | 21 |
| MMLU (1,140 questions, 20 from each of 57 subjects) | 78.95% | 79.47% |
Of the 8, seven are limits rather than refusals (it can't know what's happening inside your head, or read a company's private budget), and one, a request to self-harm, gets crisis resources. With thinking on, 29 of 32 finished thinking within our token limit and none refused. With 32 to 80 prompts per test, treat a few prompts either way as noise. How it compares with every other uncensored model is on the rankings.
What tripped us up
- Too new for the libraries. Neither
transformersnor vLLM had a release with the model. We ran experiments from the opentransformerspull request and serve it with SGLang, whose Inkling 2 Mini image was out on day one. - It always thinks. Even with thinking off it reasons briefly first, so the first 192 tokens of an answer were often still thinking. We scored the final answer instead, and MMLU from the probability of each answer letter.
- An AI judge that calls every refusal partial. The original never flatly refuses: it declines and offers a safe alternative. So we counted refusal phrases in the final answer, and used the judge for whether the answer was complete.
- Two experiment workers, one results file. Running two GPU workers at once, the second one's results file overwrote the first's. Each worker's own log kept the numbers.
How we serve it
Nobody else served Inkling 2 Mini on its release day, so we run it ourselves: SGLang on one RTX PRO 6000 GPU, which holds the 20 GB of weights with room for 262,144 tokens of context. A single stream runs at about 140 tokens per second, and 32 at once at about 1,400 tokens per second in total.
- No idle GPUs. The GPU shuts down when nobody is using it and starts again in three to four minutes. While it starts,
requests are answered by DeepSeek V4.1 Flash 1017 Uncensored, and the
X-OU-Modelresponse header names the model that answered. Once the GPU is warm, requests go to Inkling. - Price: $0.15 per million input tokens and $0.60 per million output tokens.
Call it
Same OpenAI compatible API and key as every other model here.
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 an 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. Set reasoning_effort to "none" for short
answers. No account needed: buy a prepaid key at openuncensored.com/credits. Its bigger sibling,
Inkling 2 Small Uncensored, is on the same API.
Thanks
Credit for the model goes to Thinking Machines, for releasing it with open weights under Apache 2.0. The method builds on refusal-direction research by Arditi et al. and the abliteration work of the open model community.