The Anonymous Model That Ate the Leaderboard: Space Bunny and China's Stealth AI Playbook

The Bunny Nobody Saw Coming

There was no launch event. No blog post. No benchmark card. No press outreach, no embargoed briefings, no founder thread on X.

On September 23, 2026, a model listed as Space Bunny Alpha quietly appeared on OpenRouter — the model-routing platform that has become the de facto scoreboard for the global AI industry — and on OpenCode, a developer-focused coding platform. The listing said it was a Flash-class model optimized for fast reasoning and programming. It offered a one-million-token context window, accepted text, image, and video inputs, supported adjustable reasoning intensity, and provided tool-calling capability.

The price was $0 per million tokens. Inbound and outbound. Free.

Nobody knew who built it. The developer field was blank.

Four days later, over the Mid-Autumn Festival holiday in China, Space Bunny did something no anonymous model had ever done: it topped the daily token-consumption rankings on both OpenRouter and OpenCode simultaneously, and stayed there for three consecutive days. On September 27 alone, developers ran approximately 3.99 trillion tokens through Space Bunny on OpenRouter and roughly 8.3 trillion on OpenCode — more than any named model from OpenAI, Google, Anthropic, or any Chinese lab with an actual logo on the door.

By the end of its first week, Space Bunny Alpha had processed 13.9 trillion tokens, placing it third in the world on OpenRouter's weekly global rankings — ahead of every model except DeepSeek's V4.1 Flash and Zhipu's GLM 5.3 Flash. Two of the three most-used AI models on Earth were now Chinese. The third had no name at all.

The AI developer internet did what it always does when confronted with a mystery: it started investigating. What followed was a week of crowdsourced forensic analysis — tokenizer fingerprinting, behavioral probing, contradiction cataloging — that pointed to one increasingly obvious suspect. On September 27, the same day Space Bunny hit its peak, MiniMax released M3.1-Flash-Preview, a coding model with strikingly similar specifications, inside its own MiniMax Code product.

MiniMax has not confirmed anything. As of this writing, the bunny's official identity remains unverified.

But the pattern is unmistakable — and it says something important about where the AI race is heading.

The Spec Sheet

Space Bunny Alpha's published specifications, as listed on OpenRouter, describe a model built for exactly one job: fast, reliable, long-context software development.

The spec sheet reads like a direct answer to what developers have been asking for all year: a model fast enough to live inside an IDE, smart enough to handle real repository-scale work, and cheap enough to use for routine tasks without thinking about the meter. The one-million-token context means an entire large codebase fits in a single session. The adjustable reasoning ladder means a developer can spend milliseconds on boilerplate and minutes on architecture.

And the price — zero — removed the last reason not to try it.

What developers did with it, within days, ranged from impressive to showy. They built web games, music players, and backend APIs. One X creator named siddharth gave Space Bunny a text prompt and a handful of images and asked for a seaside-themed game; the model produced boats, stations, and a coherent color palette, and the demo video passed 80,000 views. Another developer reportedly generated a detailed 3D model from reference images in about an hour. A third built a working invoicing application from photographs of paper receipts in four minutes.

The free window was listed as one week. That was enough.

The Takeover

To understand why an anonymous model topping the charts matters, it helps to look at the company it kept.

For the week of September 21–27, 2026, OpenRouter's global weekly token rankings told a remarkable story: the three most-used models in the world were all Chinese, and the third one did not have a name.

Space Bunny did not sneak into the top ten. It arrived at number three in its first week, processing roughly 71% of the volume of the world's most-used model — a model backed by DeepSeek's full brand weight and open-weight credibility.

The daily numbers were even more stark. On September 27, Space Bunny sat at number one on both of its platforms, a double crown it held for three straight days through the Mid-Autumn holiday.

The timing was not accidental. China's Mid-Autumn Festival — one of the country's longest holidays, when engineers are idle, curious, and looking for something to play with — gave the anonymous model a perfect window. Chinese developers discovered it, Chinese developers stress-tested it, and by the time Western developers returned from their weekend, the bunny had already hopped to the top of the board.

Zoom out further, and the structural picture becomes even clearer. China has now led the world in weekly AI model token consumption for 22 consecutive weeks — nearly half a year of uninterrupted dominance.

China's weekly volume of 62.22 trillion tokens is roughly 4.4 times the American total. Even on a down week — China dipped 7.77% while the global total grew 13.18% — the gap is not close. Space Bunny's 13.9 trillion tokens alone would nearly equal the entire US ecosystem.

The Detective Work

An anonymous model at the top of the global leaderboard is, in the AI community, an irresistible provocation. Within hours of Space Bunny's appearance, developers across Reddit, X, and Chinese forums began the collective work of unmasking it.

The first and most obvious technique was the one everyone uses: ask the model who it is. The results were a study in contradiction.

Some users reported Space Bunny identifying itself in ways that hinted at Moonshot AI's Kimi line — the natural first guess, since Moonshot had been telegraphing a next-generation model. Others got responses that referenced OpenAI, which complicated the picture. A few got answers that mentioned nothing at all, just the kind of careful deflection that suggests either a well-tuned system prompt or genuine uncertainty baked into the model's identity layer.

Direct interrogation was a dead end. So the community turned to forensics.

The decisive technique was tokenizer fingerprinting. Every large language model breaks text into tokens using a proprietary tokenizer, and each tokenizer produces slightly different token counts for the same strings — a signature as distinctive as a fingerprint. A third-party test ran 50 test strings through Space Bunny Alpha's tokenizer and compared the output against every major model family.

The result: Space Bunny Alpha's token counts matched MiniMax's family exactly, across all 50 strings.

A perfect 50-for-50 match does not happen by coincidence. Tokenizers are trained on a model's specific training data and vocabulary decisions; two models share a tokenizer fingerprint only when they come from the same lineage. The community's working conclusion: Space Bunny Alpha is a stealth preview of MiniMax's next Flash model, released anonymously to gather real-world feedback at massive scale without the reputational risk of a formal launch.

Reddit developers went further, speculating that Space Bunny was specifically the preview build of what became M3.1-Flash — a theory that gained credibility when MiniMax shipped that exact model four days later.

MiniMax, for its part, has maintained complete silence on the question. The bunny remains officially anonymous.

The Suspect

Whether or not MiniMax ever confirms it, the company's behavior during the same week tells its own story.

On September 27 — the day Space Bunny hit its three-day, double-platform winning streak — MiniMax quietly added M3.1-Flash-Preview to the model picker inside MiniMax Code, its AI coding assistant. No announcement came first. Developers noticed the new option in the product before the company said anything. When MiniMax finally acknowledged it, the confirmation came through a promotional post on X, not a formal launch.

The contrast with MiniMax's own history is striking. In June 2026, the company launched M3 with full architecture details, benchmark scores, and a pricing sheet. This time: no model card, no price, no benchmark, no public API identifier. The only way to use M3.1-Flash-Preview is inside MiniMax Code.

The specifications line up almost perfectly with Space Bunny Alpha: one-million-token context, multimodal input, adjustable reasoning up to a "max" tier, coding-first positioning. The only differences are cosmetic — M3.1-Flash-Preview currently outputs text only, while Space Bunny's listing mentions the same modality set.

MiniMax accompanied the release with a promotional campaign: from September 28 to October 7, users who check in daily on MiniMax Code receive double free credits, usable across M3.1-Flash-Preview and the company's H3 video-generation models. Token Plan quotas were reset for all users at launch.

One detail from the same week underlines how strange this strategy is: on September 28, MiniMax's stock — it trades in Hong Kong under ticker 00100 — dropped 10.1% to HK$247.4, even as its suspected model was dominating the global conversation. The market, apparently, has not yet figured out how to price a company that wins the leaderboard anonymously.

The Playbook

Space Bunny is not the first time a Chinese model has appeared out of nowhere and let the internet do the marketing. It is, however, the most successful execution of a pattern that has become almost standard in China this year.

The pattern is consistent: release anonymously, price at zero, let developers discover it, let the community build the detective narrative, then confirm — or quietly launch the named version — once the model has already proven itself in public.

The logic is not mysterious. A formal launch invites formal scrutiny: benchmark tables that can be compared against rivals, price points that can be undercut, capability claims that can be challenged. An anonymous release invites none of that. It generates curiosity, word-of-mouth, and — critically — real-world usage data at a scale no internal testing program could match. By the time the named version ships, the company already knows what breaks, what developers actually want, and what the model is worth.

There is also a subtler benefit: plausible deniability. If the anonymous model flops, no brand is attached. If it underperforms on a specific benchmark, there is no benchmark to point to. If it behaves badly in public, the company can disavow it — as MiniMax still technically can with Space Bunny.

In China's hyper-competitive model market — where a new flagship seems to ship every week and price wars have compressed API margins to near zero — the anonymous launch has become a hedge against the brutal economics of visibility.

The Token War Behind the Bunny

Space Bunny's emergence is not an isolated event. It is the sharpest signal yet of a structural shift in how AI models compete.

For most of 2026, the global AI race has been measured less by benchmark scores and more by token throughput — the sheer volume of compute processed through a model in real-world use. It is a harder metric to game than curated evaluations, and it captures something benchmarks cannot: what developers actually choose to use, at scale, when money and time are on the line.

By that metric, China's position has become overwhelming. Chinese models have led global weekly token consumption for 22 straight weeks. The week Space Bunny entered the charts, Chinese models processed 62.22 trillion tokens — more than four times the American total — even as the overall market grew 13.18% week-over-week.

The composition of the top five tells its own story. DeepSeek's V4.1 Flash, a free open-weights model released September 10, holds the top spot with 19.6 trillion tokens and is still growing 24% week-over-week. Zhipu's GLM 5.3 Flash sits second at 16.3 trillion. Space Bunny — anonymous, unpriced, and unsponsored — outsold Tencent's Hy4 Preview nearly one-and-a-half to one. Xiaomi's newly open-sourced MiMo-V2.6-Flash entered the chart at number eight in its first week.

Notably absent from the very top: any American closed-weight model. The frontier closed models — Anthropic's Claude Fable 5.1, OpenAI's GPT-6 Astra — still win on capability benchmarks. But on the scoreboard that measures what the world actually runs at scale, the leaderboard has a distinctly Chinese accent.

Xiaomi's MiMo-V2.6 series, released open-source on September 22, adds another dimension. MiMo-V2.6-Pro, with 1.02 trillion total parameters and 42 billion active, scored 46 on the Artificial Analysis Intelligence Index — the highest of any open-source model, ahead of Moonshot's Kimi K3 and Alibaba's Qwen3.8 Max — at a price one-twentieth to one-sixtieth of overseas models at comparable intelligence levels.

The bunny did not appear in a vacuum. It appeared in the middle of an open-source, low-price, high-throughput ecosystem that has been methodically dismantling the premium-model business model for months.

What People Are Saying

The story exploded across Chinese tech social media during the Mid-Autumn holiday, with developers treating the unmasking of Space Bunny as a collaborative sport. Here is what the conversation looked like.

@量子位上班摸鱼中 (Weibo)

中秋节别家都在发月饼，MiniMax发了一只兔子，还不承认是自己家的。这个营销策略我给满分，模型本身也真的好用，写代码快得不像话。

>

*"Everyone else sent mooncakes for Mid-Autumn Festival. MiniMax sent a bunny — and won't admit it's theirs. Full marks for the marketing strategy. The model itself is genuinely good; it writes code absurdly fast."*

知乎用户 · 匿名 (Zhihu)

匿名上线免费白嫖一周，让全网开发者替你跑测试，收集真实场景数据，最后再正名发布。这套流程已经被智谱和MiniMax玩明白了，下次再看到匿名模型直接默认中国厂商就行。

>

*"Launch anonymously, let the entire internet's developers test it for free for a week, collect real-world scenario data, then officially name it at launch. Zhipu and MiniMax have this playbook figured out. Next time you see an anonymous model, just assume it's a Chinese lab."*

v2ex.com 用户 code_mao (V2EX)

测了一下午，Space Bunny 修 bug 的准确率比 Claude Fable 差一点，但考虑到它免费而且有 100 万上下文，这点差距完全可以接受。我已经把日常用的模型切过去了。

>

*"Tested it all afternoon. Space Bunny's bug-fixing accuracy is slightly below Claude Fable, but given that it's free with a 1M context window, that gap is totally acceptable. I've already switched my daily driver."*

@AI产品观察 (WeChat public account comment)

真正可怕的不是太空兔是谁，而是它上线四天就冲到全球第三。中国模型现在不需要发布会、不需要品牌、不需要 benchmark 分数，光靠产品力就能拿到全球调用量。这才是让硅谷睡不着觉的事。

>

*"What's truly frightening isn't who Space Bunny is — it's that it hit number three globally in four days. Chinese models now don't need launch events, don't need brands, don't need benchmark scores. Pure product quality alone can win global usage. That's what should keep Silicon Valley awake at night."*

@美股韭菜日记 (X/Twitter)

MiniMax股票跌了10%，但它们的匿名模型登顶全球调用榜。要么市场蠢，要么市场知道些我们不知道的事。我倾向于前者。

>

*"MiniMax stock dropped 10% while its anonymous model topped the global usage charts. Either the market is stupid, or it knows something we don't. I lean toward the former."*

@后端小刘 (Bilibili comment)

用太空兔写了个后端接口，从需求到部署花了二十分钟，中间还让它自己写了单元测试。免费模型做到这个程度，某些收我每月两百块的模型该反思一下了。

>

*"Used Space Bunny to write a backend API — from requirements to deployment in twenty minutes, and I even had it write the unit tests itself. A free model performing at this level should make some of the models charging me 200 yuan a month reconsider their life choices."*

The Last Word

Space Bunny Alpha will not stay anonymous forever. Either MiniMax confirms what the tokenizer fingerprints already suggest, or the free window closes and the model quietly retires, its data-collection mission complete.

But the bunny's brief, brilliant run already taught the industry something worth remembering. In 2026, the most powerful launch strategy in AI is not a keynote, a benchmark, or a billion-dollar marketing budget. It is a free API endpoint, a catchy name, and the confidence to let the world's developers discover the product on their own.

China's labs have learned this lesson faster than anyone. They are operating in the world's most brutally competitive model market — dozens of capable labs, razor-thin margins, a developer population that is price-sensitive, technically sophisticated, and quick to switch. In that environment, stealth is not a gimmick. It is risk management, market research, and marketing rolled into one.

The deeper signal is the one hiding in the weekly token tables. For 22 consecutive weeks, Chinese models have processed more tokens than any other country's — and the gap is not narrowing. Space Bunny did not just top a leaderboard; it demonstrated that China's AI ecosystem can produce a globally competitive model, at global scale, with no brand at all.

When your anonymous side project can out-draw every named competitor on Earth, the question is no longer whether China's models can compete. It is whether the rest of the world can keep up with a playbook it hasn't figured out how to counter.

The bunny is fast, free, and — for now — officially nobody's. The industry should get used to it. There will be more.