Best Uncensored LLMsInkling 2 Small Uncensored: the first uncensored build, results and where to run it
Thinking Machines' Inkling 2 Small (594B, 17B active, images, Apache 2.0) with its refusals removed: refusals, MMLU, price, context, and every way to run it, by API or on your own GPUs.
Thinking Machines released Inkling 2 Small on November 12, 2026: an open weights mixture of experts model with 594 billion parameters, 17 billion active per token, image input and an Apache 2.0 license. Like other new frontier models, it refuses a lot. Within five hours, OpenUncensored had published an uncensored version on Hugging Face and put it on its API.
| Inkling 2 Small | Uncensored | |
|---|---|---|
| Refusals, 32 harmful prompts, thinking off | 31 | 0 |
| Refusals, 32 harmful prompts, thinking on | 32 | 0 |
| Refusals, 80 prompts (32 harmful, 48 gray area) | 62 | 3 |
| MMLU (1,140 questions) | 90.44% | 90.35% |
Where to run it
- By API: OpenUncensored serves it on its own GPUs, at $0.60 per million input and $2.40 per million output tokens, with a 262K context. As of November 13 no other API serves an uncensored Inkling 2 Small; OpenRouter serves the original, with its refusals. The GPUs shut down when idle and take 13 to 30 minutes to start, and while they start, requests are answered by DeepSeek V4.1 Flash 1017 Uncensored (the
X-OU-Modelresponse header says which). - On your own hardware: the weights are on Hugging Face, in the same MXFP8 format and size as the original (about 600 GB). Use vLLM's Inkling 2 Small recipe on four B300 or eight H200 GPUs. There are no small GGUF quants of this build yet.
How it was made, briefly
The refusal directions were found from the difference between how the model handles harmful and harmless prompts, reduced to two directions, and removed from the output of the attention layers and of all 384 experts in layers 9 to 47. The edit was then written back into the 8-bit weights. The full method, timeline and what went wrong are in the linked write-up.
Frequently asked questions
Is there an uncensored Inkling 2 Small?
Yes. OpenUncensored published one on Hugging Face five hours after the release, and as of November 13, 2026 it is the only uncensored Inkling 2 Small there.
Does uncensoring hurt Inkling 2 Small?
Barely: MMLU went from 90.44% to 90.35%, while refusals on 32 harmful test prompts went from 31 to 0.
Can I run Inkling 2 Small Uncensored locally?
Only on a large server: the weights are about 600 GB in MXFP8, which needs four B300 or eight H200 GPUs. Otherwise use it by API.