kimi-k3-in-c
Copyright 2026 the kimi-k3-in-c authors

This product includes software developed as part of this project, licensed under the
Apache License, Version 2.0.

================================================================================
THIRD-PARTY COMPONENTS
================================================================================

The following files under third_party/ are vendored from external projects. They remain
under their original licenses and copyrights, and are licensed to this project under the
Apache License, Version 2.0.

  third_party/json.h              minimal JSON parser
  third_party/tok.h               byte-level BPE tokenizer: loader, pre-tokenizers,
                                  encode and decode
  third_party/tok_unicode.h       Unicode character-class tables (cl100k)
  third_party/tok_unicode_o200k.h Unicode character-class tables (o200k)

MODIFICATIONS

  Per section 4(b) of the Apache License, the following changes were made to the
  vendored files for use in this project:

    tok.h    Added the `kimi` pre-tokenizer family flag (a leading \p{Han}-run rule,
             Han excluded from the letter classes) and the `rankbpe` flag, which selects
             tiktoken's lowest-concatenated-id merge rule for vocabularies that ship no
             merges list. Both are required by the Kimi K3 tokenizer and neither changes
             behaviour when unset. Comments were translated to English and the file
             header rewritten to describe the file generically rather than by the model
             it was originally written for.

    json.h   Comments translated to English; header rewritten to state the inputs this
             project actually parses. No functional change.

    tok_unicode_o200k.h   Header comment corrected to name the regex family rather than
             a specific model. No functional change.

  src/tokenizer/k3_tok.h is original work, not vendored. It populates the tok.h
  structures directly from the released tiktoken.model and tokenizer_config.json, so no
  intermediate tokenizer.json conversion step is required.

================================================================================
MODEL WEIGHTS
================================================================================

This repository contains no model weights.

Kimi K3 is released by Moonshot AI under its own license. Obtain the weights from the
official distribution and comply with that license; nothing here grants any rights to
them. docs/kimi-k3-tech-report.pdf is included for reference and remains the property of
its authors.