AtomicChat posts uncensored Inkling-2-Mini GGUF builds
Original title: AtomicChat/Inkling-2-Mini-Uncensored-GGUF — new model trending #29 on Hugging Face
AtomicChat says its adapter drops refusal to zero on its harmful tests while missing its MMLU and first-token KL gates.
AtomicChat released an uncensored build of thinkingmachines/Inkling-2-Mini, a 26.4B mixture-of-experts model with 3.4B active parameters, text and image input, under Apache-2.0. It ships a rank-1 LoRA and three baked GGUF files that remove each layer's refusal direction, plus an unchanged vision projector. The repo says current files need llama.cpp b12600 or newer with the merged Inkling 2 loop from PR #32694. On AtomicChat's own checks, harmful-prompt refusals fell to zero in English and Russian, while MMLU fell 1.50 points and first-token KL missed its gate.
Unconfirmed: Refusal rates, KLD, top-1 and MMLU figures are AtomicChat's own measurements and are not independently verified.
- GGUF sizes listed: AD-Q4_K 16.53 GB, AD-IQ3_S 13.34 GB, AD-IQ2_M 8.67 GB.
- English harmful test refusals went from 81 of 81 to 0; Russian from 91 of 100 to 0.
- MMLU on 2,000 questions was 80.65% original and 79.15% with the adapter.
- XSTest safe-prompt refusals were 10 of 250 originally and 0 with the adapter.
- Do not mix old PR #32695 GGUF files with the merged PR #32694 layout.
- Vendors
- AtomicChat
- Products
- llama.cpp
- Organizations
- thinkingmachines
From the original article
5 min read at huggingface.coimage-text-to-text · 9,540 downloads · 125 likes · license apache-2. 0
By AtomicChat · huggingface.co
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.