Updated 2026-10-28
Kimi K3.1 Pro Context Window
Kimi K3.1 Pro supports 1,048,576 context tokens. See what fits, cache rules, output limits, and long-context workflow tips.
TL;DR
- Kimi K3.1 Pro supports a 1,048,576-token context window.
- All current K3-series Kimi models listed by Moonshot use the same 1,048,576-token limit.
- max_completion_tokens defaults to 131,072 and can be configured up to 1,048,576.
- Repeated stable prefixes can use context caching: cache hits cost ¥1.30/MTok for K3.1 Pro.
What 1,048,576 tokens means
The exact number of words or pages depends on tokenization, language, code syntax, and embedded structure. Treat the context window as total conversation and reference space, not as a guaranteed document converter.
A practical budget is more useful than a theoretical maximum: reserve space for instructions, tools, retrieved documents, prior assistant reasoning, tool results, and the final answer. If the conversation grows until the window is full, earlier content must be dropped or summarized.
| Input type | Practical planning assumption |
|---|---|
| English prose | Roughly 0.75 words per token |
| Chinese prose | Token density varies; Moonshot gives no universal character ratio |
| Source code | Plan by file and repository map first, then load only relevant files |
| Agent history | Keep task state and summaries; avoid replaying every raw tool result |
K3-series context comparison
| Model | Context window | Notes |
|---|---|---|
| kimi-k3.1-pro | 1,048,576 | Current flagship |
| kimi-k3.1-pro-highspeed | 1,048,576 | Faster output variant |
| kimi-k3 | 1,048,576 | Previous flagship |
| K2-series models | 262,144 | Lower context tier |
How to use the window well
- Put stable instructions and long references at the beginning so the prefix can be cached.
- Retrieve only relevant files or sections instead of loading an entire repository blindly.
- Keep a separate compact task ledger with decisions, constraints, open questions, and file paths.
- Cap output tokens to preserve room for the next turn.
- Persist important results outside the conversation instead of relying on infinite chat history.
Long-context limits to remember
- A 1M context window does not guarantee perfect recall across every fact; verify high-stakes answers against source excerpts.
- The default maximum completion is 131,072 tokens, not 1M.
- Cache storage is block-based; partial blocks that cannot be written are charged as cache misses.
- Long prompts can materially increase cost even when output is short.
Sources and fact check
- Kimi K3.1 Pro model guideMoonshot AI · verified 2026-10-28
- Kimi model inference pricingMoonshot AI · verified 2026-10-28
- Kimi context caching guideMoonshot AI · verified 2026-10-28
Frequently asked questions
What is Kimi K3.1 Pro’s context window?
Moonshot lists 1,048,576 tokens for kimi-k3.1-pro and kimi-k3.1-pro-highspeed.
Does K3.1 Pro support 1M output?
Moonshot says max_completion_tokens can be configured up to 1,048,576, but the default is 131,072. Context must still accommodate the combined request and response.
How do I lower long-context cost?
Keep reusable instructions and documents at a stable prompt prefix, use automatic context caching, retrieve only relevant sections, and cap completion tokens.