【 以下文字转载自 NewExpress 讨论区 】
发信人: anylinkin (ALK), 信区: NewExpress
标 题: Re: 关于“AI模型开源”社区论坛说不是代码开源而是权重参数开
发信站: 水木社区 (Sat Feb 1 09:37:00 2025), 站内
DeepSeek自己的回答:"DeepSeek is open-weight with inference code, but not fu
lly open-source."
并且,它还告诉我,AI model的Open-Source也是分等级的,并且直接否认仅仅开放权重
参数不等于开源,因此它认为LLaMA自称是开源的,实际不是开源。
Fully Open-Source: Requires releasing code + weights + training details/data
. Examples include Pythia (code, weights, and dataset links) and BLOOM.
Partially Open: Many models are "open" only in the sense that weights and in
ference code are public, but training code/data are withheld (e.g., Stabilit
y AI's Stable Diffusion shares weights but not full training data).
Open Weights ≠ Open Source: Projects like LLaMA (weights-only) are sometime
s mislabeled as "open-source" but lack critical components for full openness
.
不过社区的讨论都认为,DeepSeek是目前 AI领域“开源”程度最好的模型。
【 在 Jessler 的大作中提到: 】
: 你自己看看有没有code不就可以了,github code source里都有license的
--
FROM 223.104.38.*