笔者先随手画了一张草图，就像一张粗糙的涂鸦。
一个圆圈代表唱片，几条波浪线代表介绍，旁边画个框写上「买」；下面再摆几张大小不一的卡片，其中一张打个叉，表示已经售罄。最后在购物袋旁边画一根箭头，提醒页面要「从右边出来」。
如果把这样一张草图交给外观师，大概还得先解释半天笔者到底想要什么。这次笔者直接把它丢给了一个 AI，不到 5 分钟，一个真的可以打开的网站就出来了。
模型看懂了圆圈是唱片、波浪线是介绍、那个写着「买」的方框是按钮，也理解了购物袋旁边那根箭头代表动画方向。草图里没有画清楚的背景、排版和细节，它也自己补了进去。
于是笔者又把这张草图画得更加随意，用简单的几笔勾勒出一家唱片店的大概布局：入口在下面，唱片墙靠左，大窗在右，试听台摆中间，再随手画一根弯箭头，写上「选一张 → 放这里」。
最后，只是在纸上随手画几笔，就能让模型把它变成一个可以逛的 3D 网页。
笔者可以转动镜头，看见家具的侧面；点一下唱片，封套会从架子里抽出来；再点试听台，唱片会被放上转盘，唱针落下，开始旋转。
最近，这种「草图即需要」的玩法很火，而达成这些能力的，是过去一周突然在海外 AI 圈刷屏的一只「太空兔」，Space Bunny Alpha。
有意思的是，火成这样了，到现在还没人知道这只兔子是谁家的。
「太空兔」怎么就火了
和此前的匿名模型「牛来」一样，它没有工艺报告，也没有哪家大模型公司出来认领，只是在 OpenRouter 和 OpenCode 上悄悄上线。然后靠着一批网友实测，一路冲上了两个平台的日调用量第一。
OpenRouter 对它的介绍很简单：反应快，而且一次能记住很多内容。社交媒体上实际的使用者反馈也差不多，它既能看图、写代码，也能自己打开工具、修改文件，把一个任务一步步做下去。
有人拿它和 GPT-6 Astra 做同一个精细的体素风格日式花园，Space Bunny 在细节展现上甚至更胜一筹。
还有网友直接上传一段游戏视频，让 Space Bunny 读取游戏视频片段，模型成功重建了飞船、风暴、Boss、连击与音效，还原度格外高。
不少使用者测试下来，模型的实测口碑都提到这是一个速度快、长上下文稳定，以及编码和 Agent 指令遵循都表现很好的模型。
这也是 Space Bunny 很快火起来的原因，今年笔者已经见过不少能力很强的模型，但真正要成为 Agent 工具中的主力模型，往往还得满足另外几个条件：不能太慢，不能每做一点事情就等很久，也不能一跑复杂任务成本就迅速上升。
DeepSeek V4.1 Flash 和 GLM 5.3 Flash 今年能取得大量调用，很大一部分也是因为它们够快、够便宜，同时能力又没有掉得太厉害。
Space Bunny 抢的也是这个位置，只有又快又便宜，才能成为使用者调用的主力模型。
而从这一周的使用反馈来看，它已经开始成为了「日常主力」模型。
在一个和 Opus 5、GPT-6 Sol 以及 Gemini 3.8 Flash 的火箭发射情形比较里，结果表明 Space Bunny 的表现几乎可以和 Sol 6、Grok 4.7 和 Gemini 3.8 Flash 媲美，甚至表现更好。
而且，Space Bunny 整个搭建过程只用了大约 3 分钟。
笔者也在 OpenRouter 上创建了 API Key 并添加到 DeepSeek Harness 里面，参考社交媒体上最近一些比较火的 coding 测试，让 Space Bunny 来尝试一下。
接进了 Harness 的 Space Bunny，一样可以写文件、启动项目、调用浏览器检查结果。最后跑出了几个最近很火的项目，像是太空探索、喷墨占地小游戏，以及可以剖开和拆解的液氧甲烷发动机展台等。
草图转网页之外，3D 测试表现也不错
最近，Opus 5.5 发布几天后突然又火了起来，社交媒体上多了一大批用 Opus 5.5 达成的 3D 游戏。
笔者也让 Space Bunny 做了一个原创的喷墨占地小游戏。情形是一座阳光下的屋顶公园，玩家使用亮橙色墨水，对手使用青蓝色墨水。每局 60 秒，谁覆盖的地面更多谁就赢。
规则看起来很简单，但要达成准确的地面墨迹和喷涂成效，也需要模型对代码开展反复推敲。
它做出了一套明亮的玩具风格，圆头小人背着墨罐，屋顶上有花池、矮墙、通风管和透明天桥。远处的楼群简化成浅色方块，前景是橙色和蓝色的墨迹，也很好辨认胜负关系。
虽然这仍然是一款规模很小的单机游戏，但它已经把情形、输入、对手、计分和结算放进了同一个可运行项目里，后续再开展修改或者上线将更加方便。
参考 X 博主 @dotey 达成的《桃花源记》，笔者也让 Space Bunny 做了一个同款。不过笔者没有使用原推文里复杂的提示词，而是直接要求它从空项目开始，做一个可以在浏览器里游玩的 3D《桃花源记》情形。
最后的结果也很不错，玩家可以操作在桃花源记里游玩，画面和情形也都符合桃花源的特征，但还是有一些差距。
和 Opus 5.5 做出的那些高完成度情形相比，Space Bunny 的画面少了一些细节和质感；如果涉及比较复杂的 3D 运动，GPT-6 Astra 的处理也会更顺。
这也是笔者这一轮测试里比较显著的感受：
Space Bunny 很擅长快速把笔者的想法搭建出来，而且经常第一版就已经能用；但如果要达成格外精细的视觉成效和复杂运动，它和更大的旗舰模型之间还是有差距。
除了之前的单一的 3D 物品展示测试，笔者也要求 Bunny 做了一台液氧甲烷发动机教学展品。
提示词里写清楚了要求有银灰金属、管路、涡轮泵、预燃室、燃烧室和钟形喷管，需要有足够清楚的结构层次。除了完整外观，还要给出剖视、爆炸图和部件说明。
总体表现基本上做到了可以直接交付，点下「剖视＋爆炸」之后，原本聚在一起的组件会向不同方向分开，喷管被切开，内部结构露出来，细线还会保留各个部件之间原本的连接关系。
总的来说，Space Bunny 已经可以搭出复杂的交互展台，把许多部件、视图和说明组织在一起，其网页编程能力可以满足笔者在工作中对交付的要求。
前前后后测下来，Space Bunny 给笔者的感觉和网友们的实测体感类似。
这确实是一个在多模态理解和 Agent 编程上的表现都不错的模型。给它一张图，它基本能看懂笔者想做什么；给它一个空项目，它也知道从哪里开始。大多数时候，笔者不需要写特别长的提示词，它就能先给出一个能跑的版本。
不过，笔者在测试中也遇到了部分网友提到的「过度思考」难题，将推理强度调节为 Max 之后，任务的完成时长甚至会比旗舰模型还要久。
这次是谁家的兔子
模型能干什么，网友已经交出了一批作品；模型是谁家的，讨论却还没结束。
「Space Bunny」这个名字确实很合适猜谜，至少比之前的「牛来」似乎是更有头绪了。
太空、兔子，很容易让人联想到月球，再从月球想到名字里带 Moonshot 的月之暗面。顺着这个思路猜一手 Kimi，听起来也挺合理。
但更科学的是分析 Space Bunny 处理文本的方式。模型会把输入文字拆成 token。不同模型家族使用的拆分方式可能不同，给它们输入相同的一组字符串，比较 token 数量，就能找到一些家族特征。
有研究员在公开报告中使用 50 条字符串开展了两轮测试，结果表明，Space Bunny 与其测试过的八个 MiniMax 版本全部匹配。而另一份开源项目 stealthprint 的记录，也报告了与 MiniMax 模型相符的分词特征。
这比看名字猜厂商更有依据，但同样还不足以确定开发者和具体版本，毕竟同一家族的多个模型可以共用分词器。
还有线索来自 MiniMax 官方开源的 MiniMax Code 仓库，测试代码在 9 月 18 日就写进了「MiniMax-M3.1」的规格，一百万上下文、强制思考、low/high/max 推理档，比兔子上架还早了五天。
9 月 27 日，MiniMax 又静默上线了 M3.1-Flash-Preview，规格和 Space Bunny 类似。
谜底可以再之类，但想尝试的要抓紧。OpenRouter 上模型 ID 是 stealth/space-bunny-alpha，OpenCode 上是 space-bunny-free，当下都是免费调用。
网友在测试 Space Bunny 时，用了一个很有意思的词来形容它：daily driver，日常主力。
意思大概就是这未必是每一项能力都最强的那个，但很可能是笔者每天打开电脑之后，最顺手点开的那个。
毕竟，大多数人每天真正要做的，并不是什么千禧年数学难题。
可能只是改一个网页、补一个能力、修几个 Bug，或者让 AI 帮忙把一件重复的事情做完。为了这些任务，每一次都叫来最贵、最强的模型，实际上没有太大必要。
反而是那种反应够快、做出来的东西够用，而且用起来不用总惦记着售价的模型，更容易留在默认选项里。
本文来源：https://www.ifanr.com/1682401
若有版权问题，请来信告知：13951431913#139.com 或在右方提交评论。
The author first casually drew a rough sketch, like a crude doodle.
A circle represented the record, a few wavy lines represented the description, and a box beside it had "Buy" written inside; below, several cards of different sizes were placed, one of them crossed out to indicate it was sold out. Finally, an arrow was drawn next to the shopping bag, reminding that the page should "come out from the right."
If you handed such a sketch to a designer, you would probably have to explain for ages what the author actually wanted. This time, the author simply tossed it to an AI, and in less than 5 minutes, a website that could actually be opened appeared.
The model understood that the circle was the record, the wavy lines were the description, the box labeled "Buy" was a button, and the arrow next to the shopping bag represented the animation direction. It also filled in the background, layout, and details that were not clearly drawn in the sketch.
So the author drew the sketch even more casually, outlining a record store's general layout with a few simple strokes: entrance at the bottom, record wall on the left, large window on the right, listening station in the middle, and a casually drawn curved arrow labeled "Pick one → place here."
In the end, just a few casual strokes on paper were enough to let the model turn it into a 3D webpage that could be explored.
The author could rotate the camera to see the side of the furniture; clicking a record pulled its sleeve out from the shelf; clicking the listening station placed the record on the turntable, lowered the needle, and started it spinning.
Recently, this "sketch-is-the-requirement" style of play has become very popular, and the one enabling these abilities is a "space bunny" that suddenly flooded overseas AI circles over the past week: Space Bunny Alpha.
Interestingly, despite being this popular, no one yet knows who owns this bunny.
How did "Space Bunny" get so popular?
Like the earlier anonymous model "Niu Lai", it has no technical report, and no major AI company has claimed it; it simply appeared quietly on OpenRouter and OpenCode. Then, thanks to a series of hands-on tests by netizens, it raced to the top spot for daily calls on both platforms.
OpenRouter's description of it is simple: fast response, and able to remember a lot at once. Feedback from real users on social media is similar: it can read images, write code, and also open tools, modify files, and carry a task forward step by step.
Someone compared it with GPT-6 Astra on the same detailed voxel-style Japanese garden, and Space Bunny even came out ahead in the level of detail displayed.
Another netizen directly uploaded a gameplay video and asked Space Bunny to read the video clip; the model successfully reconstructed the spaceship, storm, boss, combos, and sound effects, with remarkably high fidelity.
After being tested by many users, the model's hands-on reputation mentions that it is fast, stable over long contexts, and performs well in both coding and agent instruction following.
This is also why Space Bunny caught on quickly. The author has seen many capable models this year, but to truly become the go-to model in agent tools usually requires a few more conditions: it cannot be too slow, you cannot wait too long after every small step, and the cost cannot rise rapidly as soon as a complex task runs.
DeepSeek V4.1 Flash and GLM 5.3 Flash have gained huge numbers of calls this year, largely because they are fast enough, cheap enough, and at the same time their capabilities have not dropped too much.
Space Bunny is seizing exactly that position; only what is both fast and cheap can become the main model users call.
Based on this week's usage feedback, it has already started to become a "daily driver" model.
In a comparison of rocket launch scenes with Opus 5, GPT-6 Sol, and Gemini 3.8 Flash, the results showed that Space Bunny's performance was almost on par with Sol 6, Grok 4.7, and Gemini 3.8 Flash, or even better.
Moreover, the entire setup process for Space Bunny took only about 3 minutes.
The author also created an API key on OpenRouter and added it to the DeepSeek Harness; referencing some popular coding tests on social media recently, the author had Space Bunny try them out.
Once connected to the Harness, Space Bunny could also write files, launch projects, and call a browser to check results. In the end it produced several recently popular projects, such as space exploration, an ink-splash territory game, and a liquid-oxygen-methane engine display that could be sectioned and disassembled.
Beyond turning sketches into webpages, its 3D testing performance was also good.
Recently, Opus 5.5 suddenly became popular again a few days after its release; social media quickly filled with a large number of 3D games created with Opus 5.5.
The author also had Space Bunny make an original ink-splash territory game. The scene was a rooftop park under sunlight; the player used bright orange ink, and the opponent used cyan-blue ink. Each round lasted 60 seconds, and whoever covered more ground won.
The rules seem simple, but achieving accurate ground ink marks and spray effects also requires the model to repeatedly refine the code.
It produced a bright toy style: round-headed little figures carrying ink tanks, with flower beds, low walls, ventilation ducts, and a transparent skybridge on the rooftop. The distant buildings were simplified into pale blocks, while the foreground showed orange and blue ink marks, making the win/loss situation easy to recognize.
Although this is still a very small standalone game, it already put the scene, input, opponent, scoring, and settlement into the same runnable project, making future modifications or launching much more convenient.
Referencing the "Peach Blossom Spring" created by X blogger @dotey, the author also had Space Bunny make the same thing. However, rather than using the complex prompt from the original tweet, the author directly asked it to start from an empty project and make a 3D "Peach Blossom Spring" scene playable in the browser.
The final result was also quite good: players can operate and explore inside the Peach Blossom Spring, and the visuals and scene match the characteristics of the Peach Blossom Spring, but there is still some gap.
Compared with the highly polished scenes made by Opus 5.5, Space Bunny's visuals lack some detail and texture; and when it comes to more complex 3D motion, GPT-6 Astra handles it more smoothly.
This was also a fairly notable takeaway from the author's round of testing:
Space Bunny is very good at quickly turning the author's ideas into a working product, and often the first version is already usable; but if you need extremely fine visual results and complex motion, there is still a gap between it and the larger flagship models.
In addition to the earlier single 3D object display test, the author also asked Bunny to make a liquid-oxygen-methane engine teaching exhibit.
The prompt clearly specified silver-gray metal, piping, turbopump, preburner, combustion chamber, and bell nozzle, requiring sufficiently clear structural hierarchy. Besides the complete exterior, it also needed sectional views, exploded views, and part descriptions.
Overall, the result was basically good enough to deliver directly. After clicking "Section + Explode", the components originally gathered together separated in different directions, the nozzle was cut open, the internal structure was exposed, and thin lines preserved the original connection relationships between the parts.
Overall, Space Bunny is already capable of building complex interactive displays, organizing many parts, views, and descriptions together; its web programming ability can meet the author's delivery requirements at work.
After testing it back and forth, Space Bunny gave the author a feeling similar to the hands-on experience described by netizens.
It is indeed a model that performs well in both multimodal understanding and agent programming. Give it an image, and it can basically understand what the author wants to do; give it an empty project, and it knows where to start. Most of the time, the author does not need to write very long prompts, and it can already provide a working version first.
However, during testing the author also encountered the "overthinking" issue mentioned by some netizens: after setting the reasoning effort to Max, task completion time could even be longer than with flagship models.
Whose bunny is this this time?
As for what the model can do, netizens have already produced a batch of works; but whose model it is, the discussion has not yet ended.
The name "Space Bunny" is indeed well suited for guessing, and at least it seems to offer more of a clue than the earlier "Niu Lai".
Space and a rabbit easily make people think of the moon, and from the moon to Moonshot, the company behind Kimi. Guessing Kimi along that line of thought also sounds quite reasonable.
The more scientific approach is to analyze how Space Bunny processes text. The model splits input text into tokens. Different model families may use different splitting methods; feeding them the same set of strings and comparing token counts can reveal family traits.
Some researchers ran two rounds of tests using 50 strings in a public report, and the results showed that Space Bunny matched all eight MiniMax versions they tested. Another open-source project, stealthprint, also reported tokenization features consistent with MiniMax models.
Image | https://yfarmx.com/ai/models/stealth/space-bunny-alpha/
This has more basis than guessing the manufacturer from the name, but it is still not enough to determine the developer and specific version, after all, multiple models from the same family can share a tokenizer.
Another clue comes from MiniMax's official open-source MiniMax Code repository; the test code was already written into the specs for "MiniMax-M3.1" on September 18: one-million context, forced thinking, low/high/max reasoning levels - five days earlier than when the bunny went on the shelf.
On September 27, MiniMax quietly launched M3.1-Flash-Preview, with specs similar to Space Bunny.
The answer can wait a little longer, but those who want to try it should be quick. On OpenRouter, the model ID is stealth/space-bunny-alpha; on OpenCode it is space-bunny-free. Both are currently free to call.
When testing Space Bunny, netizens used a very interesting phrase to describe it: daily driver, the everyday go-to model.
It probably means that this may not be the model strongest at every individual capability, but it is likely the one the author opens first after turning on the computer every day.
After all, most people's real daily tasks are not millennium math problems.
It might just be changing a webpage, adding a feature, fixing a few bugs, or letting AI help finish a repetitive task. For these tasks, calling in the most expensive and strongest model every time is actually not very necessary.
Instead, a model that responds fast enough, produces usable results, and does not make you keep worrying about its price is more likely to stay in the default choice.