Do You Really Need a 1M-Token Context Window?

你真的需要1M的上下文？

Do you really need a 1M context window?

现在的大语言模型，没有1M上下文都不好意思发布，甚至2M、10M的都冒出来了。宣传语都差不多：一口气读完整本巨著、整个代码仓库一次看光。

These days, a large language model without a 1M context window is almost embarrassed to launch — 2M and even 10M are showing up. The marketing copy is all the same: read an entire epic in one pass, scan a whole codebase at once.

听起来很厉害

Sounds impressive.

但你有没有遇到过：明明塞的东西远没到1M，它的回答质量却越来越差？要么漏掉关键信息，要么前后矛盾，要么干脆开始一本正经地胡说八道。关键好像同样的用，Token越花越多。想想半年前256k上下文的模型你做多少工作量，你话了多少Token。而现在拿点Token你能做些什么？！

But have you noticed: even when you're nowhere near 1M tokens, the quality of its answers keeps dropping? Either it misses key information, contradicts itself, or simply starts confidently making things up. And somehow, for the same amount of work, you're spending more and more tokens. Think about what you could do with a 256k-context model half a year ago, and how many tokens it cost. Now — what can you even do with the same token budget?!

这不是你的错觉。而且我可以直接告诉你结论：对绝大多数人来说，1M上下文不仅用不上，反而是个坑。

It's not your imagination. And I can give you the conclusion right away: for the vast majority of people, a 1M context isn't just unnecessary — it's a trap.

先澄清一个常见误会：1M 上下文，不等于它能记住你。

First, let's clear up a common misconception: a 1M context does not mean it remembers you.

很多人以为上下文大了，模型就能记住你的喜好、你的习惯、你上次说过什么——这不是一回事。上下文窗口解决的是"这一次对话里，它能同时看多少东西"。关掉这个对话，窗口里的东西就没了。

Many people assume a bigger context means the model remembers your preferences, your habits, what you said last time — that's not how it works. A context window determines "how much it can look at at once within a single conversation." Close that conversation, and everything in the window is gone.

真正的"记住你"，靠的是另一套系统：把你的关键信息（喜好、习惯、做过的事）提炼出来、单独存好、每次对话时按需调取回来。这叫记忆系统，是Harness的设计，跟上下文大小无关。

Real "memory of you" comes from a different system: extracting your key information (preferences, habits, things you've done), storing it separately, and retrieving it on demand in each conversation. That's a memory system — a harness-level design — and it has nothing to do with context window size.

那在上下文中我们最应该关注的是什么？

So what should we actually care about in context?

回答是：模型注意力，但模型注意力是有限的。

The answer: model attention — and model attention is limited.

你可以把大模型的注意力想象成一盏聚光灯，光的亮度是固定的。照一个小房间，每个角落都亮；照一个大广场，每个地方都暗下去。

Think of a model's attention as a spotlight with fixed brightness. Shine it on a small room and every corner is lit; shine it on a large square and everywhere goes dim.

而且这不只是比喻，是真有数学约束的：模型给所有内容分配的关注，加起来永远是百分之百。塞进去的东西越多，每条分到的就越少，真正重要的信息就越容易被淹没。这个现象有个名字，叫"上下文腐烂"：输入越长，输出越差。

And this isn't just a metaphor — it's a mathematical constraint: the attention a model allocates to all content always sums to one hundred percent. The more you stuff in, the less each item gets, and the more easily truly important information gets drowned out. This phenomenon has a name: "context rot" — the longer the input, the worse the output.

既然注意力会被稀释，那自然就存在一个发挥最好的区间：长度在这个区间内，模型火力全开；超出区间，性能开始下坡。

Since attention gets diluted, there's naturally a range where the model performs at its best: within it, the model fires on all cylinders; beyond it, performance slides downhill.

这个区间，就是甜区。

That range is the sweet spot.

那甜区到底在哪？

So where exactly is the sweet spot?

今年很多模型基准测试做了大量实测（MRCR、RULER、HELMET 这几个基准，分别测多文档检索、真实上下文利用率和下游任务），结论出奇一致：

This year, many model benchmarks ran extensive tests (MRCR, RULER, HELMET — covering multi-document retrieval, real context utilization, and downstream tasks respectively), and the conclusions are strikingly consistent:

50K以内，差不多几页到十几页文档，所有主流模型都发挥正常，几乎没有差别。这个区间是日常使用的绝对主力：聊个天、写个东西、查个资料，根本到不了上限。

Within 50K tokens — roughly a few to a dozen pages — every mainstream model performs normally, with almost no difference. This range handles the bulk of daily use: chatting, writing, looking things up never even approach the limit.

50-130K，大概五到十五份长文档，或者一本中等厚度的书，这是真正的通用甜区。多份文档综合、多轮对话、跨章节推理，性能都不打折。

50–130K — about five to fifteen long documents, or a medium-length book — is the real general-purpose sweet spot. Multi-document synthesis, long multi-turn conversations, cross-chapter reasoning: performance doesn't degrade.

130-260K，开始有人掉队。只有少数旗舰还能保持不错的检索能力，但也仅限于检索。

130–260K: some models start falling behind. Only a few flagships keep decent retrieval ability — and only retrieval.

260K以上，基本只剩一种能力可靠：在超长文本里找一条具体信息，相当于给模型开了个加强版 Ctrl+F。让它综合多份文档做推理？所有模型都会明显变差，实测里有些旗舰的准确率直接腰斩。

Above 260K, basically one capability remains reliable: finding a specific piece of information in a very long text — an upgraded Ctrl+F. Ask it to reason across multiple documents? Every model visibly degrades; in testing, some flagships' accuracy literally gets cut in half.

更有意思的是，各家"声称的窗口"和"实际能用的窗口"之间，差距常年稳定在百分之三十到五十。声称1M的模型，多任务实测里往往200-400K 就开始崩。厂商宣传里的1M，和你能用的1M，从来不是一回事。

More interestingly, the gap between "advertised windows" and "actually usable windows" has stayed steady at 30–50% year after year. Models advertising 1M routinely start falling apart between 200–400K in multi-task testing. The 1M in the marketing and the 1M you can actually use have never been the same number.

那 1M 到底有什么用？

So what is 1M actually good for?

也不是一无是处，它的真实价值只剩两个场景：一是大海捞针:在法律文档、财报里精确找某一条条款，这种单点检索确实能干到九成以上；二是通读式摘要:把一整本书读一遍，讲个大概。

It's not entirely useless — its real value boils down to two scenarios. One is needle-in-a-haystack: pinpointing a specific clause in legal documents or financial reports; single-point retrieval genuinely works above 90%. The other is read-through summarization: reading an entire book and giving the gist.

但请注意，这两种都是"读"，不是"想"。你让它读1M token的合同，再综合判断哪几份条款有冲突？对不起，甜区之外，想不动。

But note: both are "reading," not "thinking." Ask it to read a 1M-token contract and reason across clauses for conflicts? Sorry — outside the sweet spot, thinking doesn't work.

最极端的例子是 Meta 的 Llama 4 Scout：声称 10M 上下文，找信息确实能找到，但实测里它对120K内容的阅读理解，准确率只有一成半——同一份测试，谷歌的模型能拿九成。

The most extreme example is Meta's Llama 4 Scout: it advertises a 10M context and can indeed find information — but in testing, its reading comprehension at 120K scores only 15%. On the same test, Google's model scores 90%.

10M的"窗户"，120K的"光线"。

A 10M window, with 120K of light.

聊天越长，忘得越狠

The longer the chat, the harder it forgets

还有一个跟日常关系更密切的：对话场景里，上下文越长，模型对早期内容的记忆越差。

Here's something closer to daily life: in conversations, the longer the context, the worse the model remembers early content.

你的聊天历史大部分是寒暄、修正、中间过程——对当前问题毫无帮助，却实实在在消耗着注意力。你以为它"记得"你前面说过什么，其实它只记得最近那几轮。

Most of your chat history is pleasantries, corrections, and intermediate steps — useless for the current question, yet consuming real attention. You think it "remembers" what you said earlier; in fact it only remembers the last few turns.

有团队实测过：同样一个问题，只给300个token的精炼上下文，和塞110K的完整聊天记录，正确率差出一大截。

A team tested this: the same question, given a distilled 300-token context versus a full 110K chat log — the accuracy gap is huge.

所以当你感觉它开始"忘事"，别怀疑，你的直觉是对的。这时候果断开新对话，把关键信息重新交代一遍，比硬撑着聊强得多。

So when you feel it starting to "forget," don't doubt yourself — your instinct is right. Start a new conversation and restate the key information; it beats pushing through the old one.

普通用户该怎么办？

What should ordinary users do?

我的建议很简单，四条：

My advice is simple — four points:

日常对话、写作、查资料，50K 完全够用。别为"窗口大"多付钱——窗口越大，单次请求越贵，对话越长，每轮都更贵。

For daily chat, writing, and research, 50K is more than enough. Don't pay extra for a "big window" — bigger windows cost more per request, and longer conversations cost more per turn.

要处理长文档，128K 的主流旗舰完全够用。除非你要一次性通读整本中文长篇小说，那是极少数场景，不在此列。

For long documents, a 128K flagship is plenty. Unless you need to read an entire Chinese novel in one pass — that's a rare scenario, not in this list.

感觉它忘事，就果断压缩或开新对话，硬撑着聊，质量只会越来越差。

When it seems to forget, compress or start fresh — pushing on only makes quality worse.

需要综合大量资料时，别一股脑全塞进去。先筛出最相关的几份再给模型，效果通常比全塞好——这就是"检索增强"思路的日常版，不花一分钱，立竿见影。

When synthesizing lots of material, don't stuff everything in at once. Filter for the most relevant pieces first — results are usually better than dumping it all. This is the everyday version of "retrieval-augmented generation": free, and immediately effective.

一句话总结：厂商比的是谁家窗户大，但你要的是谁家窗户里光线好。2026年的实测已经说得很清楚：好光线只存在于前130K里，剩下的一百万，是给"找东西"准备的，不是给"想事情"准备的。

In one sentence: vendors compete over who has the bigger window, but what you want is who has better light inside. The 2026 tests are clear: good light only exists in the first 130K. The remaining million is for "finding things" — not for "thinking about things."