Unconfirmed. Leaked test data in a since-deleted SGLang pull request references 'moonshotai/Kimi-K3.1', describing it as using Kimi K3's tokenizer and chat format with a new 'medium' thinking effort. Moonshot AI has not announced the model.
Leaked code, not an official source. A commit titled 'feat(sgl-router): Kimi-K3.1 support for the K3 chat encoder', from a since-deleted SGLang pull request, adds test data for 'moonshotai/Kimi-K3.1'. Its message says Kimi-K3.1 'ships K3's tiktoken.model byte for byte' with 'the same 16 special tokens', and that the encoder's 'only semantic change is _VALID_THINKING_EFFORTS gaining "medium"' (low, medium, high, max). The K3.1 test fixtures include image-input cases. GitHub's event log shows the pull request was closed on Oct 20 about a minute after being renamed from 'Kimi-K3.1 support' to 'Kimi-K3 support'.
We record what sources show, with dates. We don't predict or score how likely a rumor is.