Nemotron 3 Super Uncensored Hits 92% MMLU

This is a fun local-model data point, but also a reminder that benchmark jumps after "uncensoring" can come from prompting, formatting, or reasoning-mode differences as much as from actual capability gains.

• –NVIDIA's official Nemotron 3 Super is already a strong open-weight hybrid MoE reasoning model, so a community ablated fork can plausibly change benchmark behavior without changing the backbone.
• –The reported 92% vs 86% gap is notable, but it is still self-reported until someone else reruns the same eval with identical settings and template handling.
• –Mac-only MLX support matters here: a 46 GB 120B-class model that runs locally on Apple Silicon is unusually practical for power users.
• –The bigger takeaway is that "safety" layers, chat templates, and reasoning toggles can materially affect multiple-choice scores, which makes benchmark interpretation tricky.