Space Bunny: who built the anonymous AI model, and how good is it
Hundreds of thousands used it for free without knowing its maker. How the maker was traced, how it compares with Opus and GPT, and what hardware it needs.
On September 23, 2026, a new entry appeared in the model list of OpenCode: Space Bunny. OpenCode is a coding assistant app. It has no model of its own: it connects over the internet to models made by others, and its developers maintain the list of available ones. That day they added a new model to the list and announced it on X. The model was free, had a one-million-token context window, and could read images and video. Nobody said who had made it. At the same time it was opened on OpenRouter, the largest marketplace where hundreds of models are sold through a single API key. Six days later it had more users in OpenCode than any other model.
A month on, the story has an ending: the maker, the price and the internals are all public. There are also reviews claiming the newcomer beat Claude Opus and GPT. This piece goes through it in order: what happened, how the maker was traced, what the benchmarks really show, and whether the model can run on hardware you own.
A free model with no maker
The OpenCode team’s post on X was one line: “Space Bunny (stealth model) is free for the next week.” The post carried three tags: 1M context, multimodal, zero data retention. On OpenRouter it was listed as Space Bunny Alpha at zero cost for input and output. Its page said the model was developed and operated by a third party that chose to stay anonymous during the preview, and that the platform only forwarded requests.
On OpenRouter this is a familiar format. Such releases are called stealth models. A lab ships a future product under a made-up name, free of charge, to test it under real load and collect feedback before the official launch. Users get a strong model for nothing, and the developer gets millions of real requests.
Space Bunny took off harder than usual. According to OpenCode statistics, by September 29 it had about 201,000 unique users and 12.3 million sessions. In that time it processed 45 trillion tokens, against 35 trillion for DeepSeek V4.1 Flash on the same platform. On OpenRouter, by the platform’s own account, it was the number-one model for ten straight days and handled 56.7 trillion tokens, more than any stealth model before it.
The model was tuned while the test was running. OpenRouter cites three figures: time to first token dropped from 1.95 to 1.22 seconds, throughput rose from 70 to 100 tokens per second, and the error rate fell from 1.4% to 0.7%.
How the maker was traced
The developer was named nowhere, but every language model has one trait it cannot hide. That trait is the tokenizer.
A model does not read text letter by letter. Before processing, text is cut into pieces called tokens: sometimes a whole word, sometimes a syllable or a couple of characters. Each lab has its own cutting rules, chosen before training starts. A finished model cannot swap its tokenizer without being trained again. Meanwhile the platform reports the token count in every response, because that count is what gets billed.
YFarmX, a site that tracks new models, used exactly that. It ran 50 strings through Space Bunny and compared the token counts with known tokenizers. All 50 of 50 matched the family of the Chinese company MiniMax: eight of its models, from MiniMax-01 to M3, split text the same way. The closest outsider was Moonshot’s Kimi K2, with 29 matches out of 50. YFarmX noted that the test identifies a family, not a specific version.
There were other clues. The public repository of MiniMax Code had already leaked the name “MiniMax-M3.1” with a one-million-token context and the same reasoning-depth levels. Space Bunny’s feature set matched the company’s previous model, M3: a million tokens of context with text, image and video input. And on September 27 MiniMax announced M3.1 Flash Preview to its subscribers without a word about Space Bunny.
How it ended
On October 5, Space Bunny Alpha vanished from OpenRouter. On October 19, MiniMax released M3.1 Flash and wrote on its blog: “Space Bunny 🐰, which processed more than 200 trillion tokens while operating anonymously, was an early version of MiniMax M3.1 Flash.” The same day the model returned to OpenRouter under its real name, free for the first 72 hours.
MiniMax cites more than 200 trillion tokens over the whole anonymous run, while OpenRouter reports 56.7 trillion on its own platform. The company has not published a breakdown for the other platforms. It has not explained the bunny name either.
MiniMax is based in Shanghai and is known for its M-series models and for video and voice generation.
What is inside
According to the model card on Hugging Face, M3.1 Flash has about 428 billion parameters. For scale, models that run on a home computer with a single graphics card usually have between 8 and 35 billion.
Only 23 billion of those 428 billion are active for each token. The design is called a mixture of experts. The model consists of many narrow specialists, and at each step a router calls only the ones it needs. That explains the main property of M3.1 Flash: it holds the knowledge of a very large model but computes like a small one. This is why it is fast and cheap.
The rest in brief. The model was trained from the start to understand text, images and video, and it outputs text. The context is up to one million tokens: several thick books or a large codebase fit into one request. Reasoning depth can be set to one of five levels.
How strong it is
MiniMax published a benchmark table. These are the company’s own numbers, and they should be read in full, not by the first row.
| Benchmark | What it tests | M3.1 Flash | Claude Opus 5 | Other models |
|---|---|---|---|---|
| Automation-Bench v1.0.6 | workflow automation | 62.8 | 50.3 | DeepSeek V4.1 Flash-1017: 60.5 |
| DeepSWE v1.1 | fixing real coding tasks | 73.7 | 74.0 | DeepSeek V4.1 Flash-1017: 77.0; GPT-5.6 Sol: 73.0 |
| BankerToolBench | finance tasks with tools | 76.2 | 76.2 | no data |
| Terminal Bench 4.0 | working in a terminal | 38.1 | 51.8 | DeepSeek V4.1 Flash-1017: 42.3 |
| OSWorld 2.0 | operating a computer | 61.7 | 70.6 | GPT-5.6 Sol: 65.7 |
| Humanity’s Last Exam | hard knowledge questions | 43.4 | 55.6 | DeepSeek V4.1 Flash: 39.2 |
| CyberGym | cybersecurity tasks | 82.2 | no data | GLM-5.3: 84.5; Kimi K3: 80.0 |
The “beat Opus and GPT” headlines come from the first two rows. In automation, M3.1 Flash really is well ahead of Opus 5. In coding it trails Opus 5 by three tenths of a point and leads GPT-5.6 Sol by seven tenths, so all three are level. In finance tasks it is a tie.
Further down the picture changes. In terminal work, computer use and hard knowledge questions, Opus 5 leads by 9 to 14 points. The comparison is also with Opus 5, while Anthropic already ships the newer Opus 5.5. And in coding, the main discipline for models of this kind, the top spot in MiniMax’s own table goes to DeepSeek V4.1 Flash.
The sober reading: M3.1 Flash did not overtake the flagships, but it stands next to them on part of the tasks at a price dozens of times lower. That is its real achievement.
What independent tests measured
Artificial Analysis, a lab that runs its own model tests, adds three corrections.
Speed. MiniMax claims more than 200 tokens per second with a median above 170. Artificial Analysis measured 125, and OpenRouter showed about 100 during the stealth run.
Confident errors. In the AA-Omniscience test, 93% of the model’s wrong answers were confident errors, not refusals or partial answers. When M3.1 Flash does not know something, it invents an answer more often than it admits the gap.
Verbosity. Across the full Artificial Analysis test suite the model spent 161 million output tokens, against a median of 104 million for others. Tokens are what gets billed, so the wordiness eats part of the price advantage.
What it costs and who uses it
MiniMax API pricing is $0.15 per million input tokens and $0.60 per million output tokens, with cached input at $0.015. The company says agent tasks cost 70% less than on the previous model, M3.
On OpenRouter the model is served by 12 providers, including Fireworks, Together, SambaNova, Nebius and DigitalOcean. Most charge the same price, and the cheapest asks $0.1125 and $0.45. There are that many sellers because the weights are open and any company can host the model on its own servers.
Demand followed quickly. Here is the weekly OpenRouter ranking by tokens processed, with data through November 5:
| Rank | Model | Tokens per week | Week over week |
|---|---|---|---|
| 1 | DeepSeek V4.1 Flash 1017 | 56.5T | +33% |
| 7 | Claude Opus 5.5 | 6.53T | −17% |
| 8 | MiniMax M3.1 Flash | 6.04T | +46% |
| 9 | GPT-5.6 Luna | 4.99T | +123% |
The former Space Bunny sits right behind Opus 5.5 and is growing faster. OpenRouter itself warns how to read these figures: the ranking counts tokens, not quality, users or spend, and verbose models collect more. KRAKOBIT covered how DeepSeek took more than half of that market separately.
Can you run it yourself
The M3.1 Flash weights have been on Hugging Face since October 23 and were downloaded more than 41,000 times in a month. But “open weights” here does not mean “runs at home.”
First, the license. It is not a free license like Apache but MiniMax’s own Community License. Personal and non-commercial use is unrestricted. A commercial product has to display “Built with MiniMax M3.1-Flash” prominently and notify the company once by email, and a product earning more than $20 million a year needs separate written permission. Military use is prohibited.
Then the size. MiniMax gives no memory requirements and points to the server engines vLLM and SGLang. For desktop tools such as llama.cpp, LM Studio and Ollama, Unsloth has published compressed builds:
| Compression | File size |
|---|---|
| 1-bit | 129 GB |
| 2-bit | 135–144 GB |
| 3-bit | 160–194 GB |
| 4-bit | 249 GB |
For the model to fit entirely in memory, system RAM and video memory together must slightly exceed the file size, plus headroom for context. Even the most compressed build does not fit into a 128 GB machine. A realistic minimum is a workstation or a Mac Studio with 192 to 256 GB, and the 4-bit build needs about 300 GB. The harder the compression, the more quality is lost, and Unsloth gives no loss figures for this particular model.
There are workarounds. Enthusiasts build trimmed versions for 128 GB mini PCs by cutting part of the experts out of the model, which makes it not quite the same model. One report describes a run on four linked NVIDIA DGX Spark units. There are also engines that keep only the most used experts in memory and read the rest from disk. KRAKOBIT explained how that works using DeepSeek and an RTX 3090 as the example. Nobody has published speed measurements for M3.1 Flash on home hardware yet.
Open weights matter for a different reason. Companies and providers take them, host the model on their own servers and compete on price. That is why the model has twelve sellers and a price below the official one. For an ordinary user, API access costs less than any hardware.
Who it suits
M3.1 Flash is strong where there is a lot of repetitive work and price matters: AI agents, automation, processing large volumes of text and code, tasks with video input. A million tokens of context at 15 cents per million input tokens were built for exactly that.
The weak spots are just as clear. On precise factual questions it makes confident mistakes, so its factual answers need checking. In terminal work and computer use the flagships are clearly stronger. And for anyone choosing a model to run at home, it is simply too big.
The Space Bunny story showed one more thing: the anonymity of such models lasts days, not weeks. In OpenRouter’s daily ranking the next model without a maker, Crouton Alpha, is already climbing.