量化精度越高不一定越好：Qwen3 8B 27B 实测 Q8 与 Q4 互有胜负
Super interesting test by @atomic_chat_hq
Super interesting test by @atomic_chat_hq
@atomic_chat_hq 做了一个非常有趣的测试
Shows that more quantization precision does not automatically produce the better result.
表明更高的量化精度并不自动带来更好的结果。
Here, Qwen 3.8 27B Q8 (8-bit quantized version of the model) sometimes matched or beat Q4 (4-bit quantized) on the same voxel tasks.
在这里，Qwen 3.8 27B Q8（模型的8位量化版本）在相同的体素任务上有时与Q4（4位量化）持平或超越。
They finally recommend AD-Q5_K_M as their preferred quant, since it runs on a 32GB MacBook Air with 32K context and retains 97.3% next-token agreement with BF16.
他们最终推荐AD-Q5_K_M作为首选量化方案，因为它能在32GB MacBook Air上以32K上下文运行，并且与BF16相比保留了97.3%的下一词元一致性。
For the model weights alone:
仅就模型权重而言：
Q8_0: 28.9GB AD-Q4_K_M: 17.1GB
Q8_0: 28.9GB AD-Q4_K_M: 17.1GB
So Q4 saves 11.8GB of RAM/VRAM, or about 40.8% versus Q8.
因此，Q4相比Q8节省了11.8GB的RAM/VRAM，约节省40.8%。
更进一步：量化金融体系
看懂新闻只是起点——沿量化金融路径，把它变成能交付的工程能力