Kimi K3.1 leak: 3D generation cost is about 1/14 of GPT-6 Astra

Delivering industrial-grade interactive 3D assets in 15 minutes, rewriting the "STEM-student rules" of visual large models

English text is being prepared.

English text is being prepared.

Is Kimi K3.1 coming? This time it arrives with a very strong visual effect, and some netizens pointed out that the official video was made with Kimi K3.1.

Caption: Sharp-eyed netizens found the highlight "Kimi K3.1" in the footageAI blogger Chris not only confirmed the upcoming release, but also said he had already tested it for a week. He revealed that Kimi K3.1 performs excellently on the open-source 3D graphics library Three.js, and that the model costs only 70% of Kimi K3, with Token efficiency upgraded across the board. From the test footage Chris posted, Kimi K3.1 can be seen running a complete 24-hour day-night environment simulation, from the diffusion of morning light to the fall of night, with frame-level precision in light and shadow transitions, environmental details flowing in sync with the light, and the whole process smooth without breaks. You should know that programs written in Three.js are very resource-intensive, because they need to calculate vertices, lighting and textures in 3D space in real time. And Kimi has essentially achieved rendering-grade image quality with pure code.

01

Foreshadowing All Along: From Anonymous Leaderboard Entries to an Emerging Identity In fact, more details about Kimi K3.1 had already leaked out. At the end of September, developers found the model identifier "kimi-k3-1" in the configuration files of the Moonshot AI API platform. On October 21, the K3.1 option quietly appeared in the model switching list of Kimi Code, equivalent to the official team opening a small-scale test for developers in advance. In late October, two mysterious models with extremely strong combat power suddenly appeared in the arena, codenamed Kibitz and Kalmia. Among them, Kibitz performed strongly in front-end and 3D code tasks. In rendering tasks that test spatial imagination, such as "a pelican riding a bicycle", it relied on excellent geometric and physical intuition to generate near-perfect 3D dynamic images and vector graphics; in other blind tests of 3D voxel rendering, it won again with precise control over ultra-long instructions. It is worth noting that when generating 3D code of the same complexity, its Token consumption is only a few tenths of that of top foreign models. Soon developers, through tokenizer feature checks, inferred that Kibitz was Kimi K3.1. This fully matches Kimi's previous naming pattern: K2.5 was called Kiwido, K2.6 was called Kiwire, K3 was called Kivine, and K3.1 corresponds to Kibitz. Subsequently, a merge request in the SGLang codebase accidentally leaked more details: K3.1 added a new "Medium thinking" tier, expanding thinking intensity from three levels to four, allowing flexible allocation of compute according to task difficulty and balancing speed and depth.

02

Cases are circulating: Is this a 3D beast? In Arena blind tests, Kimi K3.1 inherited and further amplified its advantages in web front-end and graphics rendering, showing a degree of dominance in six subfields including 3D graphics (Three.js/WebGL), interactive large screens, and H5 dynamic simulation. Quite a few first-hand internal beta cases have already leaked online, and netizens exclaimed: "This is a 3D beast." To test a large model's visual and spatial reasoning abilities, there is an unavoidable test in the AI world: "a pelican riding a bicycle." This prompt looks simple but is full of pitfalls. Because this scene cannot reference ready-made pelican-on-a-bicycle material, the model must simultaneously understand the pelican's anatomical structure and the bicycle's mechanical logic, then combine two completely unrelated concepts and calculate the geometric positions purely in code. This tests the model's "concept combination ability," "physical intuition ability," and "spatial geometry ability." Caption: The messy "pelican riding a bicycle" produced by early models in actual tests. The bicycle drawn by early Claude 3.5 was just two circles plus a triangle, and the pelican was just a round ball with a banana-shaped beak, making it a veritable "representative work of abstract art." Kimi K3.1's actual test result, by contrast, maxes out the difficulty. Caption: The "pelican riding a bicycle" drawn by early Claude 3.5. In netizens' actual test cases, Kimi K3.1 created an absurd stack of a bicycle riding a pelican, yet it conforms to certain physical laws: the center of gravity holds, and the bicycle body stands stable. Besides the pelican, a crescent moon hangs in the night sky, distant hill outlines are set against the orange-red glow of the setting sun, a red-and-white striped lighthouse on the left casts a beam of light, and a small house in the background glows with warm light. In the details, ocean waves, a small crab, and the silhouette of a sailboat in the distance are all present. The overall composition is complete, the layers of light and shadow are rich, and the spatial relationships of all objects are completely correct. More importantly, this is not a pixel-generated image; it is vector rendering calculated line by line by the model in code, like perfectly recreating the entire fairy-tale scene with mathematical formulas. Netizens said bluntly: "This kind of stacked composition + light and shadow layers + understanding of object relationships is no longer something simple SVG can casually fake. The level of completion is directly comparable to Opus 5.5, and the front-end visual capabilities of domestic models have already charged into the first tier." Some netizens directly used Kimi K3.1 to generate a medieval castle, and the effect was stunning. The castle's detailed design is very rich: the main body consists of layered multi-story towers, with flags inserted at the spires and slightly raised in the wind; the castle gate drawbridge spans the moat, surrounded by mountains, grasslands, and scattered houses. There are also birds and floating clouds in the sky, and the light is compressed into the last warm color of dusk, falling on the stone walls to form transitions between light and dark, making the whole castle seem to "float in the air as the day draws to a close." Even more rare is that the floating rock mass, layered towers, attached stairs, and vegetation all sit within the same spatial logic and are self-consistent with one another. However, after testing it, this netizen said it "feels a bit like GPT-6.1 Sol Max." Others pointed out that Kimi K3.1 has added advanced control features, including time control and rendering quality switching, which can make scene effects very rich. Another case is generating an entire train of operational steam locomotives. Kimi K3.1 not only delivered extremely fast but also with a high degree of completion. In particular, it can also "take apart" the entire train for you to see: the locomotive, carriages, black cylindrical gas tanks, and red wheels—all components are disassembled into independent parts suspended at staggered heights in the frame, with each part complete in form and precise in position, like a moving "engineering drawing." From the perspective of 3D interactive engineering, Kimi K3.1 has found a new breakthrough in capability. Most mainstream overseas models remain at static modeling and basic scene rendering, while Kimi K3.1 links structural design, mechanical linkage, and interaction logic coherently together, capable of complete exploded views, dynamic speed control, and real-time data tracking. These show that the model understands physical engineering principles rather than merely piling up visuals. It retains GPT-style general understanding while also possessing Anthropic-style deep logical reasoning, which may be exactly a unique path for the iterative development of domestic AI technology. Some netizens used Kimi K3.1 to complete a high-precision, interactive 3D helicopter model in 15 minutes. Experts say these are valuable 3D assets that can greatly compress the prototyping cycle. Given Kimi K3.1's outstanding visual performance, netizens couldn't resist comparing it one by one with the world's top visual models. In a comparison with GPT-6 Astra, both models built structures in Minecraft—one built a robot-shaped building, the other a spherical building—and they seemed to have their own strengths, but Kimi K3.1's cost was only about 1/14 that of GPT-6 Astra. Caption: GPT-6 Astra and Kimi K3.1 challenge building structures in Minecraft. In a comparison with Haiku 5.5, both models built village models at the same time, both including feature settings such as dragon attacks, thunderstorms, and day-night cycles. Both took 15 minutes to generate, and with detail effects roughly on par, many netizens said they preferred Kimi K3.1's overall visual effect. However, in a real-world test titled "Which model can catch up with Claude Opus 5.5," Kimi K3.1 exposed a clear weakness: the running giraffe footage it generated had movements that did not follow normal biological motion trajectories, making it look very strange. Overall, Kimi K3.1 has already proven that Chinese large models, in the multimodal intersection of "code + vision + reasoning," possess strength that is not inferior to top US closed-source models—and in certain practical scenarios even surpasses them. This is expected to significantly lower the barrier for developing 3D and Agent applications across industries, which is good news for developers building creative applications, enterprises building visualization projects, and teams building automated workflows.

03

Kimi's visual ambition: carving out a new path through code. The reason Kimi's visual capabilities are so formidable is that it has taken a path completely different from traditional text-to-image. Simply put, traditional models "draw from impression," which easily leads to structural collapse and object clipping; Kimi first writes rigorous 3D graphics code, then renders it for self-checking, and upon finding "the wheel is flat" or "the color is wrong," automatically rewrites the code to correct it—equivalent to verifying calculations while drawing. This capability was not achieved overnight; Kimi has been laying the groundwork for nearly two years: In December 2024, Kimi launched the K1 visual reasoning model, introducing reinforcement learning into the visual domain for the first time, and surpassing foreign benchmarks for the first time in blind tests of physics and chemistry diagram problems; In April 2025, Kimi open-sourced the Kimi-VL series of models, based on the MoonViT architecture, with only 2.8 billion activated parameters, matching hundred-billion-level large models in MathVision mathematical vision and OCR, solving the pain point of "vision costing too much compute"; In July 2026, the 2.8-trillion-parameter Kimi K3 was released, directly taking first place globally in the Arena front-end code arena, surpassing top closed-source models in six major front-end domains including data visualization and interactive animation, establishing dominance in Web 3D and graphics generation; Now with K3.1, judging from its performance in anonymous tests, its advantages in 3D and front-end scenarios are expected to continue. On one hand, Kimi, at less than one-tenth the cost, has caught up with or even partially surpassed top overseas closed-source models in high-value practical scenarios such as 3D interaction and front-end engineering. On the other hand, it will focus on improving the Agent's ability to execute tasks continuously over long periods. To this end, the team plans to further optimize the model using post-training, computer operation, and real task data. Meanwhile, according to a Nomura Securities report, Moonshot AI is developing a larger model expected to have 5 to 6 trillion parameters (referred to externally as Kimi K4). Compared with Kimi K3 released in July this year, the new model's parameter scale is nearly double. However, they are currently facing the biggest obstacle of insufficient compute, and the specific training progress is not yet clear. Reference link: https://x.com/Kimi_Moonshot/status/2113445757185606062