Context Length Guide
Overview
Kimi-Researcher has a context length of 128K tokens (about 60,000–100,000 Chinese characters). This is the maximum number of tokens the model can process in a single run, including both input and output.
This means Kimi-Researcher can refer to a large amount of text in a single research task, enough to support complex and in-depth research analysis. However, please note: the maximum length of generated content is usually much smaller than the context window.
If your research question is too broad, we recommend breaking it into several sub-questions and researching them separately instead of submitting everything at once. This usually improves both depth and accuracy.
Key Concepts
| Concept | Meaning | Notes |
|---|---|---|
| Context window | The maximum token limit supported by the model | 128K tokens, including input and output |
| Input limit | The length of reference materials + instructions that can be sent in one request | Recommended to keep within 100K tokens |
| Output limit | The maximum length the model can generate in one response | Usually about 8K–16K tokens, much smaller than the context window |
Common misconception: A 128K context does not mean the model can output 128K tokens at once. Output length is usually around 1/8 to 1/16 of the context window.
Output Truncation
Why does the output sometimes stop before completion?
When a research report is too long, the model may:
- Stop proactively — stop generating after reaching the single-response output limit
- Suggest continuation — ask whether you want it to continue generating the remaining sections
- Output in sections — split a long report into multiple chapters and present them one by one
This is not a malfunction. It is normal behavior caused by output limits.
What can you do?
- If the report is incomplete, reply directly with “Continue” or “Please finish the remaining sections”
- For complex research, proactively ask for chapter-by-chapter generation, for example: “Write Part 1 first: background analysis”
Best Practices
Optimize information placement
When working with long documents, place key information at the beginning and end of the prompt rather than in the middle. The model may extract information from the middle of a long context less accurately (the "Lost in the Middle" effect).
Input strategy for long documents
- Summarize very long reference materials first instead of pasting the full text
- Mark key passages in multiple documents to reduce irrelevant noise
- Use segmented processing or retrieval-augmented strategies instead of filling the context all at once
Conversation management
- In multi-turn conversations, watch the accumulated history and start a new conversation or summarize the current one when needed
- After switching topics, start a new conversation to avoid context confusion
- Regularly summarize confirmed conclusions so that later outputs are based on accurate background information
Model selection
Choose a suitable model for your scenario and balance context length with response efficiency. A longer context is not always better—overly long context increases processing latency and computing cost.
Notes
- Context length determines how much information the model can “see”; it does not determine how much content it can “say”
- Prioritize effective delivery of key information and plan the content structure carefully
- For very long research tasks, proactively split them into sections instead of relying on a single generation