Qwen 3.5 uncensored tops 89% MMLU

This looks less like a clean intelligence jump and more like a benchmark boost from removing refusals, but it is still a serious engineering feat to make a model this large usable on local Mac hardware.

• –The key technical story is JANG mixed-precision quantization, which keeps a 397B model runnable in 112 GB while still hitting useful token speeds.
• –The 1.9-point MMLU lift over the base JANG_1L build is real on paper, but uncensoring can inflate multiple-choice benchmarks without improving downstream coding or tool use as much.
• –MLX Studio is doing a lot of the heavy lifting here; this is a very Apple-Silicon-specific win, not a general-purpose deployment recipe.
• –For local model enthusiasts, the value is obvious: more capability per watt and per GB, with multimodal support still intact.
• –For everyone else, the safer read is "better benchmark score on a derestricted variant," not "new frontier model breakthrough."