Leaks Point to Kimi K3.1's Low-Cost 3D Generation Push
Leaks point to Kimi K3.1, a code-driven 3D model reportedly costing about one-fourteenth of GPT-6 Astra.
The report said Moonshot released a teaser with rapid cuts of scientific imagery and particle effects, and users spotted the string “Kimi K3.1” in the footage. AI blogger Chris said he had tested the model for a week and that it performed well on the open-source 3D library Three.js, with cost at 70% of Kimi K3 and improved token efficiency. In his test, the model generated a 24-hour day-night environment with frame-level lighting transitions and no visible breaks, according to screenshots cited by Leiphone.
Signals of the update appeared before the teaser. The report said developers found a “kimi-k3-1” identifier in Moonshot's API platform configuration in late September, and the Kimi Code model selector added a K3.1 option on Oct. 21. In late October, two mystery models, Kibitz and Kalmia, appeared on an Arena leaderboard; Kibitz performed strongly in front-end and 3D code tasks, and developers used tokenizer characteristics to infer that it was K3.1, the report said. A merge request in the SGLang codebase also indicated a new “Medium” thinking tier, expanding the model's reasoning settings from three levels to four, according to the report.
Test cases cited by Leiphone focused on code-generated scenes rather than pixel images. The report said K3.1 produced a pelican-and-bicycle scene with correct spatial relationships and layered lighting, a medieval castle with towers, a drawbridge and surrounding terrain, and a steam train whose locomotive, cars, tanks and wheels could be separated into independent parts. A tester generated an interactive 3D helicopter in 15 minutes, the report said.
The report compared K3.1 with overseas models on cost and output. In a Minecraft building test against GPT-6 Astra, K3.1's cost was about one-fourteenth of the rival's, according to the report. In a village-building comparison with Haiku 5.5, both models took 15 minutes and some users cited by the report said they preferred K3.1's visual result. The report also described a clear failure: a running giraffe generated by K3.1 had motion that did not follow normal animal movement.
Moonshot's visual work has followed a code-first path, the report said. Rather than generating an image directly, the model writes 3D graphics code, renders it and then checks the result, rewriting code when geometry or color is wrong. The report traced the effort to the K1 visual reasoning model in December 2024, the open-source Kimi-VL series in April 2025, and the 2.8-trillion-parameter Kimi K3 in July 2026, which ranked first on an Arena front-end code leaderboard.
The report said K3.1 is expected to keep Moonshot's advantage in 3D and front-end tasks while improving agents' ability to execute long tasks; the team plans to use post-training, computer operation and real-task data. It also cited a Nomura Securities report saying Moonshot is developing a larger model of 5 trillion to 6 trillion parameters, referred to externally as Kimi K4, but is facing compute shortages and has not disclosed training progress.
Editor's Summary Leiphone reported that Kimi K3.1, linked to a teaser and developer leaks, has shown strong code-driven 3D and front-end generation in developer tests, with one comparison putting its cost at about one-fourteenth of GPT-6 Astra. The report also noted weaknesses in biological motion and said Moonshot is developing a larger K4 model while facing compute constraints.