Kimi token counter
Moonshot open-sourced the Kimi tokenizer: a tiktoken-style byte-pair encoder with a 163,840-token vocabulary. That makes exact, private, in-browser counting possible, and almost no token tool covers it.
About the Kimi tokenizer
The tokenizer ships with the open Kimi K2 model weights on Hugging Face. Its vocabulary is roughly 60% larger than OpenAI's o200k_base, built for heavily multilingual text and code, which means Kimi frequently spends fewer tokens on Chinese and mixed-language input than Western tokenizers do. The practical consequence: cost comparisons that assume one universal token count across providers overstate Kimi's price on multilingual workloads.
Counting and cost together
- Kimi models bill input and output separately, with cache-hit input at a large discount (a tenth of standard input on K3, about a sixth on K2.6). Current list prices are in the cross-provider pricing table.
- Context windows run from 262,144 tokens on K2.6 to 1,048,576 on K3. Compare at context windows.
- Counting methods and verification for every provider on this site: how we count.