本地大型语言模型 vs. 离线维基百科
Local LLMs versus offline Wikipedia

原始链接: https://evanhahn.com/local-llms-versus-offline-wikipedia/

最近一篇《麻省理工科技评论》的文章引发了对本地大型语言模型(LLM)和离线维基百科下载大小的比较——设想了一种在“末日”场景下,两者都可能在断网情况下发挥作用的情况。作者研究了通过Ollama提供的模型和Kiwix提供的维基百科包,重点关注在消费级硬件上可实现的尺寸。 结果显示出令人惊讶的重叠。较小的LLM,例如Qwen 3 0.6B,的大小与简易英语维基百科相当。值得注意的是,维基百科最好的5万篇文章的大小与Llama 3.2 3B大致匹配。完整的维基百科下载量甚至可以超过测试过的最大的LLM(高达57GB,而某些模型为20GB)。 然而,作者强调这并非直接比较——LLM和百科全书服务于不同的目的。文件大小并非一切;LLM需要大量的处理能力。尽管存在这些限制,但研究结果表明,大量的知识都可以以这两种格式本地存储,这使得两者都可能成为离线情况下的宝贵资源。

## 本地大语言模型 vs. 离线维基百科:总结 讨论的重点在于,在“社会重启”的情境下,哪种资源更有价值——本地大型语言模型(LLM)还是离线维基百科。维基百科提供知识的静态快照,而LLM则提供理解和适应能力。 它们可以理解表述不清晰的问题,用简单的语言解释复杂的主题,并将跨学科的思想联系起来,充当信息的向导,而不仅仅是知识的存储库。 然而,LLM容易出现不准确的情况,而维基百科尽管可能存在偏见,但通常更可靠。理想的解决方案可能是结合两者:使用LLM来优化查询,并使用离线维基百科来确保事实的准确性。 实际考虑因素包括硬件限制——旧设备可能难以运行LLM,而维基百科可以在最小的系统上运行。像Kiwix这样的项目促进了离线维基百科的访问,而RAG(检索增强生成)旨在结合两者的优势。最终,这场讨论突出了在保存知识以应对潜在未来危机时,可访问性、准确性和交互式理解之间的权衡。
相关文章

原文

Two days ago, MIT Technology review published “How to run an LLM on your laptop”. It opens with an anecdote about using offline LLMs in an apocalypse scenario. “‘It’s like having a weird, condensed, faulty version of Wikipedia, so I can help reboot society with the help of my little USB stick,’ [Simon Willison] says.”

This made me wonder: how do the sizes of local LLMs compare to the size of offline Wikipedia downloads?

I compared some models from the Ollama library to various downloads on Kiwix. I chose models that could be run on some consumer-grade hardware, and Wikipedia bundles that didn’t have images for a better comparison. Here’s what I found, ordered by size:

NameDownload size
Best of Wikipedia (best 50K articles, no details)356.9MB
Simple English Wikipedia (no details)417.5MB
Qwen 3 0.6B523MB
Simple English Wikipedia915.1MB
Deepseek-R1 1.5B1.1GB
Llama 3.2 1B1.3GB
Qwen 3 1.7B1.4GB
Best of Wikipedia (best 50K articles)1.93GB
Llama 3.2 3B2.0GB
Qwen 3 4B2.6GB
Deepseek-R1 8B5.2GB
Qwen 3 8B5.2GB
Gemma3n e2B5.6GB
Gemma3n e4B7.5GB
Deepseek-R1 14B9GB
Qwen 3 14B9.3GB
Wikipedia (no details)13.82GB
Mistral Small 3.2 24B15GB
Qwen 3 30B19GB
Deepseek-R1 32B20GB
Qwen 3 32B20GB
Wikipedia: top 1 million articles48.64GB
Wikipedia57.18GB

This comparison has many caveats:

  • This is an apples-to-oranges comparison. Encyclopedias and LLMs have different purposes, strengths, and weaknesses. They are fundamentally different technologies!

  • File size is not the only important detail. LLMs, even local ones, can use lots of memory and processor power. Offline Wikipedia will work better on my ancient, low-power laptop.

  • Other entries might be more useful for a specific purpose. For example, you can download a selection of Wikipedia articles about chemistry, or an LLM that’s better tuned for your hardware. (And Kiwix has lots of other things you can download, like all of Stack Overflow.)

  • I picked these entries based on vibes. Nothing rigorous about this comparison!

Despite those caveats, I thought it was interesting that Wikipedia’s best 50,000 articles are, very roughly, equivalent to Llama 3.2 3B. Or that Wikipedia can be smaller than the smallest LLM, and larger than the largest ones—at least in an offline scenario on consumer hardware.

Maybe I’ll download both, just in case.

联系我们 contact @ memedata.com