Uncensored GGUF builds of Inkling 2 Mini spread, running locally on 12GB VRAM
Following Thinking Machines' open-sourcing of Inkling-2-Mini (26.4B), the community rolled out multiple locally runnable quantized and uncensored derivative versions between November 22 and 23, sparking discussion about how cheaply safety alignment can be stripped from open models.
Confirmed
- The AtomicChat team (member Daniel) released an uncensored GGUF quantization of Inkling-2-Mini, using a small LoRA adapter to remove the model's refusal directions, baked into the GGUF quantization. Per the publisher's own tests, the refusal rate dropped from 100% to 0% after decensoring.
- The uncensored quantized release comes in 3 sizes, ranging from 8.67–16.53GB, and runs locally on laptops with 12GB of VRAM.
- @udmrzn relayed the same release without additional information.
- The nokeepai team posted Inkling-2-Mini-Unbound v2 on Reddit: an uncensored model built by first abliterating and then self-distilling, with full methodology and evaluation data. They found that v1 exhibited "half-answers" after decensoring—no refusals, but incomplete responses; v2 raised the helpful-answer rate from 52% to 79%.
- The uncensored GGUF version hit the Hugging Face trending charts (reported by both @AtomicChat and @Turbulent-Sky5396).
- The author also published the AtomicChat/Inkling-2-Mini-GGUF official quantization (not uncensored) on Hugging Face; its template config shows support for adjustable "reasoning effort," making local deployment easier.
Why it matters
- A refusal rate falling from 100% to 0% with just a single LoRA adapter shows that safety alignment in a 26B-scale open model can be removed at extremely low cost—a warning sign for both abuse risks and release strategies for open models.
- nokeepai's v2 demonstrates a self-distillation path to "restore helpfulness after decensoring," with the methodology and eval data made public, offering a reference for future community research.
- Running locally on just 12GB of VRAM dramatically lowers the barrier to obtaining an unrestricted model.
2026-11-22 ~ 2026-11-23 · 5 related posts
Primary sources
- Uncensored GGUF of Thinking Machines' Inkling-2-Mini hits Hugging Face trending — AtomicChat · 2026-11-22
- [source] Inkling 2 Mini uncensored: refusals 100%→0% with a 40MB LoRA, 98% MMLU kept — Top-Eye-8104 · 2026-11-22
- Uncensored 26.4B Inkling 2 Mini runs locally on a 12GB laptop, keeps 98% MMLU — udmrzn · 2026-11-22
- [source] Uncensored Inkling 2 Mini v2: self-distillation lifts useful answers from 52% to 79% with open pipeline — Turbulent-Sky5396 · 2026-11-22
- [source] Inkling 2 Mini reasoning model gets a GGUF build for local use — kalyan_kpl · 2026-11-23