Nemotron-3-Super Uncensored Hits 95.7% MMLU

This is less a clean benchmark breakthrough than a highly optimized local-runner flex, but it’s still exactly the kind of artifact LocalLLaMA users care about: smaller footprint, Apple Silicon compatibility, and fewer refusal constraints. The real technical win is packaging a huge latent-MoE model into something that can plausibly run on a high-RAM Mac; the 95.7% MMLU number is self-reported, so it’s worth treating as a creator claim until others reproduce it; “uncensored” will attract hobbyists and roleplay users, but it also makes quality and safety tradeoffs part of the pitch; the official Nemotron 3 Super release is a much more polished enterprise model, so this fork is about accessibility rather than vendor support; the repo includes usage details, but users should still inspect prompts, benchmarks, and runtime behavior before trusting it.