An independent developer built a browser-based Endurance in 5 hours with MiniMax M3.1 Flash
An independent developer used only one model, MiniMax M3.1 Flash, and worked continuously for 5 hours to bring the Endurance ring spacecraft from Interstellar into the browser, achieving real-time drag-to-rotate, interior cabin scenes, part disassembly and reattachment, and a space docking shot. The model supports million-level long context and image/video understanding, with official pricing at $0.6 per million output tokens and $0.015 for cache reads. The author believes the key is not the model's ceiling, but breaking down the task by setting acceptance criteria at three levels: long shot, medium shot, and close-up.
View original postHoly crap, last night I had MiniMax M3.1 Flash running nonstop for 5 hours, and I finally got the Endurance from Interstellar built into the browser. 12 main cabin segments arranged in a ring, warm white shell, dark gray skeleton, blue planet behind. Watching it slowly rotate is genuinely satisfying. There are close-ups of the cabin segments, interior concepts, structural disassembly and restoration, and spacecraft docking. From afar you see the outline; zoom in and you can see seams, pipes, and equipment. The biggest takeaway from finishing this project is that prompts have to be written clearly. For example, I wrote 3 levels for the visuals: Long shot: ring proportions, cabin segment rhythm, central node, and overall outline are clear. Medium shot: can distinguish functional differences among habitation modules, experiment modules, energy modules, communication modules, etc. Close-up: can see panel seams, grooves, porthole frames, hatches, handrails, heat sinks, antennas, pipes, and docking rings. Also, more specific details improve the quality feel—materials, lighting, camera presets, and interaction acceptance criteria all need to be written clearly. In the final output, the parts can return to place after being unfolded, the camera movement and composition stay stable, the spacecraft aligns with the interface, and it all has to run in real time in the browser. The whole project relies on MiniMax-M3.1-Flash. It supports million-token context, image and video understanding, and is strong at coding and multi-step execution, which is what made continuous refinement possible. Seeing the final product, I feel these 5 hours were very worth it. And according to official API pricing, output tokens cost only $0.6 per million, and cache reads only $0.015, making repeated refinement more cost-effective. Previously, this kind of work required crossing hurdles in modeling, materials, animation, and frontend. Now, with the help of models, an individual can string these together. The sci-fi scenes I loved as a kid can now be built with my own hands, which is quite satisfying.
View original
卧槽，昨晚让 MiniMax M3.1 Flash 连轴转了 5 个小时，终于把《星际穿越》的永恒号搓进浏览器了。 12 个主舱段围成一圈，暖白外壳、深灰骨架，背后蓝色行星。看着它慢慢转起来，真的有点爽。 舱段近景、舱内概念、结构分解复原、飞船对接都有。远看有轮廓，拉近能看到接缝、管线和设备。 做完这个作品最大的感受是提示词得写明白。 比如画面我写了 3 个层级 远景：环体比例、舱段节奏、中央节点和整体轮廓清楚。 中景：能够分辨居住舱、实验舱、能源舱、通信舱等功能差异。 近景：能够看到面板接缝、凹槽、舷窗边框、舱门、扶手、散热片、天线、管线和对接环。 另外越具体的细节能提高出品质感如材质、光照、镜头预设和交互的验收标准都得写清楚 最终输出的成品零件展开后能归位，镜头移动构图稳住，飞船对得上接口，还得在浏览器里实时跑。 整个项目依赖 MiniMax-M3.1-Flash，它支持百万上下文，支持图片视频理解，编程和多步骤执行也强，才有能力 持续打磨。看到最终产品感觉这 5 个小时也是非常值。 而且按官方 API 定价，每百万输出 Token 才 0.6 美元，缓存读取只要 0.015 美元，反复打磨变得更有性比价了。 以前这种作品，建模材质动画前端样样都得跨坎。 现在个人借助模型，也能把这些串起来了。 小时候喜欢的科幻场景，现在能亲手搓出来，还挺满足的。
Body · AI translation
We are truly lucky to live in an era of technological equality!! In the past, if someone wanted to build the ring-shaped Endurance ship from Interstellar in a web page, they had to overcome four major hurdles at once: 3D modeling, material texturing, physical animation, and front-end rendering. Now, an independent developer used only one model, MiniMax M3.1 Flash, worked continuously for 5 hours, and directly moved the entire ship into the browser.
Clicking on this 89-second real-machine screen recording, the visual completeness is indeed impressive. Behind it is a huge blue planet, and in front is a ring-shaped ship formed by 12 main cabin segments, with a warm white shell paired with a dark gray skeleton. Not only can it be dragged and rotated in real time in the browser and switched into conceptual interior cabin scenes,
but the control panels on the left and right sides can also be used to disassemble parts one by one and fit them back precisely, and it even fully runs through a space docking shot.
Many people's first reaction after watching is to sigh that writing 3D front-end is so easy now, but if you have ever used a large model to write Three.js yourself, you will know that when you throw in a sentence like "make me an Interstellar ship," 99% of the time what comes out is a pile of geometric blocks whose coordinates do not even line up.
Why was he able to spend 5 hours grinding out industrial-grade details? The core is that he did not treat AI as a wishing well, but directly set up an extremely strict three-layer construction acceptance standard for the model:
→ The long shot checks the overall proportions and ring silhouette, ensuring that the ship's first-impression spatial scale against the blue planet background does not distort → The medium shot checks the functional differences among the 12 main cabin segments, refusing to copy and paste the same block to muddle through → The close shot drills directly into seams, portholes, and pipeline details, while strictly specifying material and lighting texture, that disassembled parts must be able to return to their original positions, that camera push-pull must be stable, and that docking interfaces must fit precisely,
What is even more stunning is the model foundation supporting these 5 hours of high-intensity rework. For a large 3D project like this, with thousands of lines of code and dozens of rounds of local detail changes, if it were run on an expensive overseas top-tier flagship, the API bill alone would scare off individual developers,
while MiniMax M3.1 Flash not only withstood million-level long context and image and video understanding, but according to official pricing, every million output tokens costs only $0.6, and cache reads are pushed down to $0.015, directly smashing the trial-and-error cost of repeated modifications in long chains to the floor.
This is like the old days of shooting sci-fi blockbusters: if you hired a VFX master who charged by the second, you simply did not dare ask him for a second revision; now here comes a top craftsman with extremely fine handiwork who charges only a few cents in tea money for each revision. As long as the storyboard drawing in your hand is detailed enough, he can stay up all night with you to polish out a cinematic texture. This also reveals the bottom line to all friends who manage teams, lead projects, or use AI to build products themselves: many people complain that what large models write looks like a toy, but in fact it is not that the model's engineering ceiling is insufficient; rather, the person giving instructions has not even figured out how to accept long shots, medium shots, and close shots. People who can break down task levels and set acceptance standards will not be worse at leading human teams than at leading an AI legion. https://t.co/ShcWyLJHuc