Ask HN:对 GPT-5.4 等模型的实际规模/架构有什么靠谱的猜测吗?
1 分•作者: dsrtslnd23•5 个月前
有人对 GPT-5.4、Gemini 3.1 和 Opus 4.6 这样的大模型到底有多大,以及它们与 GLM-5 这样的最佳开源模型相比如何,有靠谱的直觉或确凿的线索吗?
它们现在都大致在同一量级吗(例如,大约 1 万亿参数,可能采用 MoE 架构),还是闭源模型仍然大得多?
也对“Pro”版本(如 GPT-5.4 Pro)感到好奇——这很可能是一个不同的模型,还是主要是在推理时使用更多算力 / 更长的推理链 / 更好的编排?
查看原文
Does anyone have decent intuitions or hard clues on how big models like GPT-5.4, Gemini 3.1, and Opus 4.6 actually are, and how they compare to the best open models like GLM-5?<p>Are they all roughly in the same range now (for example around 1T params, maybe MoE), or are the closed models still much bigger?<p>Also curious about “pro” versions like GPT-5.4 Pro - is that likely a different model, or mostly the same model with more inference-time compute / longer reasoning / better orchestration?