AtomicChat drops uncensored Inkling-2-Mini GGUF
AtomicChat released an uncensored build and GGUF quantization of Thinking Machines Lab's Inkling-2-Mini, removing refusal directions from the residual stream to run unrestricted on 16GB consumer hardware. The multimodal looped MoE model executes 3.4B active parameters per token while rivaling heavier dense models like Qwen3.8-27B.
Removing refusal vectors from looped MoE architectures highlights how lightweight, weight-reusing models can achieve uncensored local inference without degrading base intelligence.
- –Looped execution reuses layers 6–17 three times, delivering 48-layer reasoning depth from 26.4B total parameters while only activating ~3.4B parameters per step.
- –Residual stream ablation removes refusal vectors directly rather than relying on fine-tuning, preserving the model's base intelligence across its 1M context window.
- –Fits comfortably within 16GB unified memory via GGUF quantization, making high-end multimodal chat and vision tasks viable on entry-level MacBooks and mid-tier GPUs.
- –Requires bleeding-edge llama.cpp builds supporting draft PR #32695 to handle looped tensor graphs, signaling a minor initial barrier for turnkey local deployment.
DISCOVERED
10m ago
2026-11-20
PUBLISHED
36m ago
2026-11-20
RELEVANCE
AUTHOR
0x0SojalSec