we benchmark models nobody actually runs
英文标题：we benchmark models nobody actually runs
qwen3.8-27b looks genuinely impressive on the benchmark tables - beating models many times its size on some of them. but those numbers come from bf16 weights, and nobody here is running a 27b at bf16. we're running the 4-bit at ~17gb because that's what fits on a 4090 or a 24gb mac. so the model tha
打开原文相关 2 · 机会 0
中文
qwen3.8-27b looks genuinely impressive on the benchmark tables - beating models many times its size on some of them. but those numbers come from bf16 weights, and nobody here is running a 27b at bf16. we're running the 4-bit at ~17gb because that's what fits on a 4090 or a 24gb mac. so the model tha
中文全文翻译正在补齐，可先查看下方英文原文。
English / 原文
qwen3.8-27b looks genuinely impressive on the benchmark tables - beating models many times its size on some of them. but those numbers come from bf16 weights, and nobody here is running a 27b at bf16. we're running the 4-bit at ~17gb because that's what fits on a 4090 or a 24gb mac. so the model tha