Kimi K3 has a 1,048,576 tokens-token context window — a clean 1M, with no long-context mode to enable, no beta header, and no premium tier for using all of it. Output is configurable separately, and unusually generously: 131,072 tokens by default, up to the full 1,048,576 tokens if you ask.
What 1M tokens actually holds
| Content | Roughly |
|---|---|
| Plain English text | ~750,000 words, or ~750 pages |
| Source code | A large repository with tests and docs |
| PDFs | A few hundred to a few thousand pages, depending on layout |
| Images | Many, but they consume tokens fast at high resolution |
| Meeting transcripts | Dozens of hours of conversation |
The vocabulary is 160K tokens, which affects how efficiently text compresses into that window — particularly for non-English content.
Why the cached price changes the maths
Plenty of models advertise a million tokens. What is less common is a 10x cached-input discount alongside it.
A 500,000-token corpus costs $1.50 to send at the standard $3.00 rate. Sent again with the prefix unchanged, it costs $0.15. Twenty questions against that corpus cost roughly $4.35 in input rather than $30 — and that is the difference between "we should build retrieval" and "we can just load the documents".
This is the practical argument for using a big window rather than a pipeline, and it applies specifically because caching here is automatic rather than something you configure.
Size is not the interesting part
What matters is whether quality holds across the window — whether the model connects a clause on page 40 to a contradiction on page 600.
Moonshot's architecture is explicitly built for this: Kimi Delta Attention replaces standard quadratic attention in most layers, and Attention Residuals let later layers retrieve from earlier ones. Those are long-context design choices rather than marketing.
The practical habit that verifies it at no cost: ask for exact quotations with locations. If a citation is wrong you will see it immediately, and no benchmark tells you as much about your own corpus.
The cost of a large prompt
At $3.00 per million input tokens, a full 1M-token prompt is $3.00 before the model writes a word. Three consequences:
- Cache, or you are burning money. Repeat reads at $0.30 make iteration affordable; uncached repeats do not.
- Don't fill it out of habit. Sending more context than the task needs is a direct cost with no quality guarantee.
- Consider retrieval at genuine scale. For corpora in the tens of millions of tokens, or ones that change constantly, a pipeline still wins. For everything smaller, the window is simpler.
More on Kimi K3
Start with the complete Kimi K3 guide for the overview, or go deeper:
Ready to go deeper?
Read the full Kimi K3 guideFrequently Asked Questions
How big is Kimi K3's context window?
1,048,576 tokens — a full 1M, with no separate long-context mode to enable. That is roughly 750,000 words or 750 pages of text. Maximum output is 131,072 tokens by default and can be configured up to the full 1,048,576.
Does using the full context window cost extra?
There is no long-context premium on the per-token rate, but you pay for what you send: a 1M-token prompt costs $3.00 in input tokens. Cached reads bring repeats down to $0.30 per million, which is what makes repeated querying of a large corpus practical.
Does Kimi K3 actually use the whole window well?
Its architecture is designed for it — Kimi Delta Attention in most layers, plus Attention Residuals for retrieval from earlier layers — and Moonshot publishes strong long-context results. Independent verification on your own corpus is still worth doing: ask for exact quotations with locations and check them.
