SGLang's router adds Kimi-K3 rendering ahead of any release note mentioning the model
sgl-router PR #40390 (merged 2026-09-21) adds native XTML rendering and segmented tiktoken encoding for Kimi-K3 through the pinned Dynamo libraries, applying SGLang's request semantics so routing hashes and forwarded token IDs match the engine exactly. It reuses the request adapter for reasoning controls, tools, response formats and assistant continuations, loads local `tiktoken.model` files or downloads vocabulary siblings from Hugging Face, and validates against SGLang-generated token digests rather than Dynamo self-comparisons (945 tests passing). One honest limitation is in the body: explicit null `thinking_effort` with thinking enabled falls back to worker rendering because pinned Dynamo cannot represent it. This merged the same day as companion DeepSeek V4.1 Flash rendering PRs (#40530, #40532).
Source
↳ Follow the thread