JANG Quantization Beats MLX on MiniMax

This looks like a real fix for a specific MLX failure mode: not just smaller weights, but better answers from the same local Mac hardware budget. If the numbers hold up outside the author’s harness, JANG could be one of the most useful Apple Silicon inference tools in the local-LLM stack.

• –The headline result is stark: JANG_2S scores 74% on MiniMax-M2.5 while MLX 4-bit/3-bit/2-bit cluster around 25%, which is basically random on that test set.
• –The repo frames JANG as “the GGUF equivalent for MLX,” but with models staying in GPU memory at full Metal speed, so this is both a format and runtime story.
• –The practical upside is biggest on huge local models: the post cites Qwen3.5-122B at 79% with 38 GB versus MLX 2-bit at 56.5% with 36 GB.
• –The benchmark story is promising but still self-reported, so third-party replication will matter before anyone treats this as settled evidence.
• –Even so, the product fills a clear gap for Mac users who want better coherence than uniform MLX quantization gives them.