Nemotron-3-Super-120B runs uncensored on Apple Silicon

Community uncensored releases of frontier-class models keep pace with official launches, and Nemotron's novel hybrid architecture made this a genuinely hard technical problem to solve.

• –Nemotron-Super-120B's unique three-pathway design — 40 Mamba-2 SSM layers, 40 LatentMoE layers (512 experts, top-22), and 8 attention layers — breaks standard fp16-then-quantize workflows; all surgery must happen at quantization level
• –CRACK weight surgery targets the architectural convergence point of all three pathway types, suppressing refusal behavior at the weight level rather than via prompt injection or fine-tuning
• –4-bit MLX quant achieves 43–58 tok/s on M3 Ultra 256GB, putting 120B-class local inference within reach for well-equipped Mac users
• –LM Studio silently drops 697 essential tensors and is incompatible — only MLX Studio or vMLX work correctly, a notable gotcha for the community
• –A chat template workaround introduced occasional missing closing think tags, an acknowledged tradeoff of the approach