Release and model choices
The Pro Highspeed variant is positioned for more responsive interaction, with the announcement describing two to three times the speed of Pro. Both variants are offered across the Kimi product family and API.
The official K3.1 Pro repository publishes its tiktoken vocabulary and tokenizer configuration. This site verified the base vocabulary against the existing K3 browser conversion; that verification does not extend automatically to Highspeed or to complete XTML requests.
| Model | Local counting | Open tool |
|---|---|---|
| Kimi K3.1 Pro | Exact base text | Select Kimi K3.1 Pro |
| Kimi K3.1 Pro Highspeed | Reference encoding | Select Kimi K3.1 Pro Highspeed |
How to count Kimi K3.1 Pro tokens
Counts plain text with the verified published tokenizer, including real token IDs. Chat templates, role markers, tool definitions, media inputs and provider-added formatting are outside this base-text count.
Paste text into Plain text mode, or use Chat messages to count each message body independently. Switching models preserves your input. Chat mode excludes role names, separators and provider wrappers; adding bodies does not reproduce a serialized API conversation.
Local processing does not send your prompt to a model API or generate a response. Published tokenizer assets may be loaded by the browser, while the text analysis itself runs on your device.
Request limits and usage
K3.1 Pro's published tiktoken vocabulary at revision d9e99b0792a5e93b78ea9bea62a68167e36ddd9e matches all 163,584 base ranks in this site's tested K3 browser conversion. Its configured special-token IDs and splitting rules were checked too. Highspeed remains reference-only because its separate tokenizer mapping has not been verified. Neither mode includes XTML, tools, media or API overhead.
For a production request, include the actual conversation, system instructions, tool schemas and supported media in the provider's counting workflow when available. Check response usage after generation and account for subsequent calls, cache behavior, reasoning and output separately.
- A model's advertised context window is not a guaranteed amount of pasted text.
- A faster model or a cheaper token rate does not establish a lower total bill for every task.
- Recheck the exact model version and counting method when migrating an existing prompt.
Tiktokenizer Reference Sources
Tokenizer behavior and model limits change. Verify production decisions with current provider documentation: