Kimi teases K3.1 model in short video

Kimi’s official account published the 24-second clip on Oct. 23, posting it on both X and the WeChat Channels video platform.[1][2] The video shows an old computer running Kimi Code, with a model menu on its screen displaying “Model：Kimi K3.1.”[1] IT之家 read the teaser’s clues as indicating that K3.1 was nearing release.[1] That is an interpretation of the teaser, not a formal release announcement: IT之家 reported that Moonshot had not provided an official launch date as of its publication.[1]

Moonshot had also not announced K3.1’s parameter scale or pricing as of IT之家’s report.[1] Those missing details leave teams without confirmed information in the cited reporting to use for launch scheduling, cost estimates or capacity planning.[1] The teaser identifies the model name, but the evidence does not establish when the model will be available, what it will cost, or how large it will be.[1] Teams assessing whether to evaluate K3.1 can treat the clip as a signal to monitor, rather than as a basis for deployment decisions.[1]

An SGLang repository submission dated Oct. 19 had a title referring to Kimi K3.1 support for the K3 chat encoder.[1] The change added K3.1 chat-encoding support to an experimental routing component.[1] SGLang-related code indicates that K3.1 adds a medium reasoning-strength option to K3’s low, high and max options, while the reported default routing setting remains max.[1] IT之家 described the work as encoding-compatibility validation, not a performance benchmark.[1] The reported new tests cover conversation formats, tool calls and image-input formats.[1]

Kimi K3 was released as an open model in July.[3] IT之家 reported that K3 supports a context window of 1 million tokens and was integrated into GitHub Copilot on Aug. 6; Amazon’s AI coding assistant Kiro announced K3 integration on Oct. 15.[1] In a separate benchmark report, DeepRSI gave K3, described as a 2.8-trillion-parameter model, a score of 820 and ranked it third globally and first among Chinese and open-source models.[3] That result concerns K3, not K3.1; the reporting describes DeepRSI as a benchmark for recursive self-improvement and autonomous research capabilities.[3] TestingCatalog said K3.1 had previously been spotted in testing on Arena and had reportedly shown very good performance.[4]