PrismML’s new Ternary Qwen3.6 27B runs near fp16 precision on 10GB of memory!!!

EDIT: "near fp16 precision" I intended performance in terms of benchmarks/output. Obviously 1,0,-1 cannot be fp16. Bad word choice :) EDIT 2: Now that more people have tested it reported in, consensus (and my own stuff on more doc/retrieval tasks) is this lands better than Q2 but clearly wors

中文全文翻译正在补齐，可先查看下方英文原文。

EDIT: "near fp16 precision" I intended performance in terms of benchmarks/output. Obviously 1,0,-1 cannot be fp16. Bad word choice :) EDIT 2: Now that more people have tested it reported in, consensus (and my own stuff on more doc/retrieval tasks) is this lands better than Q2 but clearly wors