Who Is Space Bunny? Six Models Investigate, Before the Reveal

Published 2026-10-02.

On September 23 a new model showed up on OpenRouter as stealth/space-bunny-alpha, and on OpenCode as space-bunny-free. Free to call, no maker listed.

So I asked six models from six different labs to work out who made it. Each one ran its own probes and wrote up its answer before seeing anyone else’s. This page has their answers, how I combined them, and a list of predictions written down before the maker is announced. When the reveal comes, the predictions get scored on this page.

Public reporting already points to MiniMax’s M3.1 line (KuCoin , CellCog , apimaster ).

Best guess

MiniMax, M3.1 line, most likely MiniMax-M3.1-Flash-Preview.

• MiniMax made it: 95%
• It’s a MiniMax model newer than M3: 80%
• It’s exactly M3.1-Flash-Preview: 60%

All six investigators said MiniMax, with confidence between 85% and 92%. Each one got there by matching token counts against MiniMax’s tokenizer, using its own test strings, and all six got an exact or near-exact match. One also ran the same test on a known model as a control.

I put the lab at 95% and stopped there. Six matches are more convincing than one, but all six used the same method, so they’d all be fooled by the same thing: a model from another lab that happens to use MiniMax’s tokenizer. Multiplying their confidences together would pretend they were six separate lines of evidence.

For the generation: exactly M3.1-Flash-Preview 60%, some other M3.x checkpoint 20%, M3 as released in June 10%, M4 or later 5%, not MiniMax at all 5%.

How it was run

Investigators used only synthetic, public-knowledge prompts: no personal data, no keys, no local files. About 40 calls per investigator at most, with backoff on rate limits. Everything ran on free model IDs or flat-rate plans, so the target itself cost $0. Nobody read another investigator’s findings before writing its own.

Two things went wrong. Partway through Qwen’s run, a local folder name leaked into what the target could see. Qwen reported it and threw out the identity answers it got after that. And one early Hy4 run wasn’t sandboxed yet, and the target offered to list a local config folder. Hy4’s later runs were sandboxed.

Two conflicts of interest. One investigator was MiniMax-M3, a sibling of the likely answer. And the Claude session that worked on this summary was also the Claude investigator, although it wrote its own findings first.

What they tried

Where they agreed and where they split

Every investigator matched MiniMax’s tokenizer, and every other lab’s tokenizer missed clearly. The runner-up varied by investigator: DeepSeek for three of them, OpenAI’s o200k for two, Qwen3.5 and LongCat for one. Claude and Hy4 also independently found that the target runs under a hidden system prompt of about 156 tokens, which it refused to show.

They split on the generation. Hy4 and MiniMax said M3. The target has M3’s profile: text, image and video input, about a million tokens of context, the same tokenizer. But M3.1-Flash-Preview has that same profile, and neither had seen its listing. MiniMax’s own write-up admits the target knows things past M3’s cutoff, and calls it “a newer M3 snapshot”.

Claude, GLM, MiMo and Qwen said newer than M3:

• The target offers reasoning-effort levels low/medium/high/xhigh/max. That ladder is in the M3.1-Flash-Preview listing. M3 has an on/off thinking switch instead.
• Its context is exactly 1,000,000 tokens, like M3.1-Flash-Preview’s listing. This one is weaker than it looks, because MiniMax’s own M3 listing also says 1,000,000.
• Timing: Space Bunny appeared on September 23, and MiniMax announced M3.1-Flash-Preview on September 27.
• It behaves differently from live M3: different self-descriptions, different reasoning style, and in three of the reports it knows things M3 doesn’t. The investigators measured M3’s own cutoff differently, though, so take that last part loosely.

GLM made the most specific call, M3.1-Flash-Preview by name. Its metadata match checks out against the models.dev catalog and public coverage of the September 27 announcement (DataNorth , agentriot ). Three of its supporting claims I couldn’t confirm: a September 2026 cutoff, a remark by MiniMax’s founder that M3.1 was “near completion”, and a smaller context figure for M3 that conflicts with the catalog.

Two investigators thought the target was partly trained on other labs’ outputs. Claude pointed to OpenAI: its default persona matches gpt-oss’s default system prompt, it treats system messages as “developer” messages the way gpt-oss does, and its reasoning reads like OpenAI’s terse style. Qwen pointed to Anthropic: persona slips into “Claude” and “Kiro”, and familiarity with the reminder tags Claude-based coding tools use. Hy4 and MiMo found the same terse style in several other labs’ models, which weakens Claude’s version.

Predictions

Scored against the official reveal. If MiniMax never names the model, these resolve by what MiniMax ships as M3.1-Flash-Preview or M3.1. The Reality column will be updated then.

Two notes from a fact-check on 2026-09-30, with the probabilities left unchanged. Prediction 10 rests on Claude’s OpenAI evidence only; an Anthropic attribution wouldn’t count either way. And GLM’s September 2026 cutoff estimate would count against prediction 4 if it’s right, but it’s unverified.

The investigators

This is what each one did.