Kimi K3.1 leak rumors: 3D code costs, four thinking tiers, and what developers should verify
According to a Leiphone report on October 23, 2026, an official Kimi teaser contained an easter egg for "Kimi K3.1"; developers also spotted a kimi-k3-1 identifier in Moonshot's API configuration and a K3.1 option added to Kimi Code's model list; an SGLang merge request was said to add a Medium thinking tier, expanding thinking strength from three tiers to four. AI blogger Chris claimed about a week of testing and a cost around 70% of K3, and the report also said its 3D cost is about 1/14 of GPT-6 Astra. All of these are rumors and single-party testing.
What changed
A Leiphone report dated October 23, 2026 said Kimi released a teaser containing an on-screen "Kimi K3.1" easter egg, interpreted as a warm-up for a new model. The report cited AI blogger Chris as saying he had tested it for about a week, that K3.1 performed strongly on Three.js, and that its cost was only about 70% of Kimi K3 with improved token efficiency; the report also showed footage of it running a 24-hour day-night environment simulation. The report also laid out several leak clues: in late September, a developer found a kimi-k3-1 model identifier in Moonshot's API platform configuration file; on October 21, a K3.1 option was added to Kimi Code's model switching list; in late October, anonymous models code-named Kibitz and Kalmia appeared in an arena, with Kibitz performing strongly on front-end and 3D code tasks, and a developer using tokenizer characteristics to infer it was K3.1. A merge request in the SGLang codebase was said to expose that K3.1 adds a Medium thinking tier, expanding thinking strength from three tiers to four. The report said that in a Minecraft building comparison against GPT-6 Astra, K3.1's cost was about 1/14 of the other, while also showing weaknesses such as unnatural motion in animation tasks like a running giraffe. The report additionally relayed a Nomura Securities report saying Moonshot is developing a larger-parameter model, unrelated to K3.1.
What it means for developers
If these leak clues eventually materialize, the direct impact on developers centers on two things. The first is cost structure. The reported figures of "about 70% of K3" and "about 1/14 of GPT-6 Astra" are relative values rather than list prices, but they point to the same thing: in code generation tasks like Three.js, WebGL, and interactive large-screen displays that require per-frame computation of vertices, lighting, and textures, the token consumption and number of calls per unit of output could be significantly lower than current main models. For 3D prototypes, visualization walls, and H5 simulation pipelines that are mainly code generation, this directly determines whether a failed generation can be retried at low cost. The second is control granularity. The report said thinking strength goes from three tiers to four with a new Medium tier, meaning the same model might switch compute tiers between simple layout tasks and complex physics-linked tasks; if scheduling can be tiered by task difficulty, long-chain agent tasks could use a low tier for routine steps and a high tier for structural reasoning, instead of billing the entire chain at the highest tier. Note that these are still rumors and single-party testing, not officially confirmed, and the actual billing basis, tier naming, and rate-limiting policies may all differ.
Integration and evaluation
The following is conditional analysis and advice, not confirmed fact. First, regarding the claim of "about 70% of K3's cost," the choice is to run a same-prompt comparison first rather than directly migrating production traffic. The precondition is that you can get access to K3.1 and the task is code-generation type; the evaluation standard is fixed as the same set of 3D scene prompts (one each for a day-night cycle, multi-object spatial relationships, and exploded part views), recording success rate, retry count, and total consumption; only if total consumption after retries is still below the existing model baseline should you consider small-scale traffic diversion. Second, regarding the "new Medium thinking tier," the choice is to explicitly specify the tier by task complexity rather than using a uniform tier. The precondition is that the interface actually exposes a tier parameter; the evaluation standard is to run comparisons separately on structural reasoning tasks (such as mechanical linkages and exploded parts) and routine layout tasks, comparing whether the output quality gap between the high tier and Medium tier is within an acceptable threshold; if quality does not noticeably degrade with the Medium tier on routine tasks, set it as the default. Third, regarding the "tokenizer-characteristic inference of the anonymous arena model Kibitz," the choice is not to equate Kibitz's leaderboard performance with K3.1's stable capabilities based on speculation. The precondition is that what you care about is long context and very long instruction control; the evaluation standard is to retest with your own long instruction set and observe where instruction following degrades, rather than only looking at public blind-test rankings. All evaluations should confirm actual billing on the vendor's and KimiSeek's pricing pages before deciding whether to scale up.
Limits and open questions
The core information in this report comes from teaser easter eggs, configuration file fields, codebase merge requests, and speculation about an anonymous arena model, none of which has been confirmed by an official release; the figures (70%, 1/14) are rumors or single-party testing, not verified benchmarks. Kimi K3.1 does not currently appear in KimiSeek's available catalog; the API configuration, Kimi Code list, and call availability mentioned in the report all point to the vendor side, and no version of it is callable on KimiSeek; the Kimi K3 in the catalog is a different version and cannot be treated as the same model. The reported weakness case (unnatural running motion for a giraffe) and the positive cases are both user testing, with unknown sample size and methodology, and the comparison target's performance has not been independently reproduced. The thinking tier details in the SGLang merge request, the actual tier names, and billing differences cannot be confirmed from this report. The Nomura Securities statement about a larger-parameter model relayed in the report is unrelated to K3.1, and it said the effort is constrained by compute with unclear progress. For actual pricing and availability, refer to the vendor's and KimiSeek's pricing pages; prices may change at any time.
Illustrations from the original report
