去哪里找到这些模型 Where to Find LLMs

我们已经讲清楚了这些模型是怎么训练出来的、它们究竟是什么。那么很自然的下一个问题是:你到底去哪里才能找到、并真正用上它们?答案取决于你想要哪一类模型,大致可以分成四种情况,这一章逐一过一遍。 We've covered how these models are trained and what they actually are. So a natural next question is: where do you actually find them, and how do you use them? The answer depends on which kind of model you want. It roughly splits into four cases, and this chapter walks through each.

第一种情况:最大、最强的专有(proprietary)模型。这一类你基本上就是直接去对应提供方自己的网站。比如想用 OpenAI 的模型,就去 ChatGPT;其他前沿提供方也各有自己的官方入口。这些是闭源的旗舰模型,只能在提供方自家的界面或 API 上访问。 Case one: the biggest, strongest proprietary models. For these you basically just go to the website of that provider. To use OpenAI's models, you go to ChatGPT; other frontier providers each have their own official entry point too. These are closed flagship models, accessible only through the provider's own interface or API.

但提供方很多、模型迭代又快,你怎么知道「此刻到底谁最强」?这就要看排行榜(leaderboard)。一个常被提到的是 LM Arena 排行榜——它通过让人类对不同模型的回答做盲测对比来排名。再搭配一个 AI 新闻聚合站(比如「AI News」这类),帮你追踪每天的新进展。排行榜回答「现在谁领先」,新闻站回答「最近发生了什么」。 But there are many providers and models iterate fast — so how do you know who is currently best? That's what a leaderboard is for. A commonly cited one is the LM Arena leaderboard, which ranks models by having humans do blind, head-to-head comparisons of their answers. Pair it with an AI-news aggregator (something like 'AI News') to track day-to-day developments. The leaderboard answers 'who leads right now,' the news site answers 'what just happened.'

第三种情况:开放权重(open-weights)模型,比如 DeepSeek、Llama。这类模型权重是公开的,但你未必想自己买 GPU 来跑,于是可以去一个推理服务商(inference provider)。常被推荐的一个是 Together.ai——进它的 playground,就能挑各种不同的开放模型,直接在上面对话。要注意:这些服务商基本都面向「助手 / 聊天」场景,所以你通常找到的是经过后训练的对话模型,而不是原始的 base 模型。 Case three: open-weights models, e.g. DeepSeek and Llama. Their weights are public, but you may not want to buy GPUs to run them yourself, so you go to an inference provider. A commonly recommended one is Together.ai — open its playground and you can pick from many different open models and talk to them directly. Note: these providers are mostly geared toward assistant / chat use, so what you usually find are post-trained conversational models, not raw base models.

第四种情况:在自己的电脑上本地(locally)运行。最大的模型(比如完整版 DeepSeek)是塞不进笔记本的,但有更小的版本:经过蒸馏(distillation)、再用更低的数值精度(precision)运行——不必跑原生精度(如 DeepSeek 的 FP8、Llama 的 BF16),而是远低于此。这样就能装进你的电脑,在笔记本上跑出还不错的效果。常用的工具是 LM Studio:一个本地应用,加载好模型后就能直接对话——全程在你自己机器的 GPU 上运行,不经过任何第三方服务器。它的界面偏专业、有点杂乱(会列出一堆其实用不上的模型,挑选不同蒸馏版本和精度时容易让人犯晕),但弄懂之后就很顺手。 Case four: running locally on your own computer. The largest models (e.g. full DeepSeek) won't fit on a laptop, but smaller versions exist: distilled, and run at lower numerical precision — not the native precision (e.g. FP8 for DeepSeek, BF16 for Llama) but far below that. That makes them fit on your machine and run quite decently on a laptop. A common tool is LM Studio: a local app where, once you load a model, you can talk to it directly — entirely on your own machine's GPU, never going to any third-party server. Its interface is professional-leaning and a bit cluttered (it lists many models you won't need, and picking among distillations and precisions can be confusing), but once you understand it, it's handy.