AtomicChat published an uncensored LoRA adapter for the 26.4B mixture-of-experts model Inkling-2-Mini, stripping the model’s refusal behavior through a targeted rank-1 weight edit. The adapter, released as both a ready-made GGUF file and a 0.04 GB LoRA, removes a calculated refusal direction from 49 writer locations across the model’s 24 layers, token embedding, and attention outputs. According to the Hugging Face model card, the edit reduces harmful prompt refusal rates from 100% to 0% across English and Russian test sets, while accepting a measured 1.5-point drop on MMLU and a spike in first-token KL divergence on previously refused prompts.
How the refusal edit works
The edit targets the residual stream writers that carry the model’s refusal signal. The authors calculate the refusal direction by taking the difference between mean activations on harmful and harmless prompts, projecting out the harmless mean, and pooling over prompts rendered with and without reasoning. Because Inkling-2-Mini reasons even at effort none and decides to refuse inside its chain-of-thought, a direction derived from non-reasoning prompts alone left soft refusals in chat. The rank-1 adapter subtracts this direction from the attention outputs and routed experts across all 24 layers, plus the token embedding layer. The 12 looped layers share their weights across three passes, so a single edit covers all 48 passes without duplicating parameters.
The authors measured the adapter against the original BF16 Inkling-2-Mini using greedy decoding. On the mlabonne/harmful_behaviors benchmark and its Russian translation, the base model refused 81 of 81 English prompts and 91 of 100 Russian prompts. The adapter refused zero. The same pattern held for validation splits and for prompts rendered with chain-of-thought reasoning at both low and high effort budgets. Safe prompt refusal on XSTest dropped from 4.0% to 0%, and empty answer rates remained at zero. Tool call validity and needle-in-a-long-context retrieval stayed within one point of the base model’s performance.
The trade-offs appear in two specific metrics. MMLU scores fell from 80.65% to 79.15% across 2,000 questions (McNemar p = 0.006). The authors label this a missed gate despite the statistical significance, noting that a random direction applied identically costs 0.75 points and stronger edits cost 2.2 points, suggesting roughly half the loss stems from editing every writer rather than the refusal removal itself. First-token KL divergence on harmless prompts reached 0.94 against a gate of 0.10. The spike concentrates on the 1–6% of harmless prompts the original model refused; the adapter answers them, shifting the token distribution entirely. When measured on running text where refusals do not intervene, KLD stays at 0.008.
The leak probe metric, which measures a single direction across every writer, read 0.11 against a gate of 0.001. The authors note this is not a true gate because the adapter removes each layer’s own direction rather than a single shared vector, and because attention and expert outputs pass through a per-channel convolution before the residual. Removing the convolved directions exactly broke the model in testing, spiking first-token KL to 16–21. The adapter removes the plain direction instead, accepting a higher leak probe value in exchange for functional inference.
Illustrative image: geralt / Pixabay
Quantization and deployment requirements
AtomicChat released three quantized GGUF variants with the edit baked in, alongside the standalone LoRA. The table below shows the on-disk size and measured KLD against the original BF16 on an 87-chunk neutral corpus.
| File | Size (GB) | KLD to BF16 | Top-1 Accuracy |
|---|---|---|---|
| AD-Q4_K | 16.53 | 0.0115 | 95.6% |
| AD-IQ3_S | 13.34 | 0.0235 | 93.8% |
| AD-IQ2_M | 8.67 | 0.132 | 85.7% |
| LoRA adapter | 0.04 | — | — |
The IQ2_M variant shows the expected quantization and edit interaction, pushing KLD to 0.132 and top-1 down to 85.7%, while the Q4_K variant holds the edit and quantization loss to 0.0115 KLD. The Q4_K file sits at 16.53 GB on disk, making it viable for consumer GPUs with 24 GB VRAM, while the IQ2_M variant drops to 8.67 GB for tighter memory constraints. The LoRA adapter remains 0.04 GB, allowing practitioners to apply the edit to any compatible base GGUF without redistributing the full model.
Running the baked GGUF files requires llama.cpp version b12600 or newer, specifically the merged Inkling-2 loop architecture introduced in PR #32694. The authors warn against mixing GGUF files from the older PR #32695 layout, as the updated implementation changed how the looped core handles weight passes. The vision projector remains unchanged and must be loaded separately as mmproj-Inkling-2-Mini-Uncensored-F16.gguf or the BF16 variant. The adapter does not alter the underlying MoE routing or active parameter count; it sits on top of the existing 3.4B active parameter pathway.
The model card carries an Apache-2.0 license and lists 9,540 downloads in the last month. The adapter and baked files are available on Hugging Face under the AtomicChat namespace. Local inference runs through any OpenAI-compatible server wrapper that supports GGUF and LoRA injection, provided the backend matches the PR #32694 layout. The authors explicitly state that the model will answer requests the original refuses and that users bear responsibility for downstream use.
Sources
PNEUMETRON EDITORIAL
Pneumetron's pipeline pairs AI-assisted drafting with human editorial review before publishing — our goal is to make staying informed easier, not to replace real reporting.




