fastokens v2 adds native support for tiktoken models, including Moonshot's Kimi K3, and replaces its pretokenization regex with a hand-written Unicode scanner while rebuilding the BPE core for both serving latency and bulk throughput. All changes are validated bit-for-bit against the HuggingFace tokenizer. The update allows loading OpenAI-style tiktoken.model files directly, so Kimi K3 and other tiktoken-only models work without format conversion. The new pretokenizer uses direct byte checks for ASCII and tables for non-ASCII, with boundaries identical to tiktoken's on English, code, CJK, and diverse input. On Kimi K3 (Apple M4 Pro), fastokens v2 is 10–46× faster than HuggingFace tokenizers across production models and 10× faster than tiktoken on inference- and training-relevant workloads, with parallelization within the document. For serving, it includes a prefix cache for shared system prompts, tokenizing a 1M-token prompt that's 90% shared in ~1 ms/request, and encode_batch_flat, which returns a batch as one compact uint32 buffer plus offsets. Future work includes additional tiktoken pattern families and further prefix cache parallelism. fastokens is open source.