I genuinely believe big ai corpos made deals behind the scenes to keep the models as bloated as physically possible. There is no reason for model weights to be in the hundreds of gb. It has been PROVEN and DEMONSTRATED that you can fit 27b models in 6.8gb size (not just model weight but everything included to run it) with ternary architecture while maintaining near 90% of the performance. There is not GOD DAMN REASON to use floating point bullshit anymore to shit up free space. If the big corpos just made everything as efficient as that then every schmuck willing to spend 1k on a graphics card with 24 gb vram could run a 100b model all on their own and there wouldn't be a dime of revenue, let alone profit, for the greedy corpos.
我真相信那些AI大厂背后做了见不得人的交易，故意让模型膨胀到物理极限。模型权重搞到几百GB完全没道理。已经有实证，三元架构能把27b模型塞进6.8GB（不光是权重，连跑起来所需要的一切都算上），还能保持近90%的性能。没有半点狗屁理由再用浮点那套玩意儿来糟蹋硬盘空间了。要是大厂把所有东西都做到那么高效，那随便哪个肯花1000刀买块24GB显存显卡的路人，都能自己跑个100b的模型，那些贪婪的大厂就别想再赚一毛钱，更别提利润了。
