Space Bunny Alpha › Identity

Is Space Bunny Alpha MiniMax?

Not confirmed — but it is the best-supported hypothesis, and there is real measurement behind it. An independent tokenizer analysis matched MiniMax on 24 of 24 probes, and a verified context ceiling of at least 1M tokens rules out the older m2.5 generation. What is still missing is the only thing that would actually settle it: a vendor claiming the model.

Last updated 25 September 2026 · Model released 23 September 2026

The short answer

Probability, not fact. The evidence points to the MiniMax model family. No company has claimed Space Bunny Alpha, TokenRA lists the provider as anonymous, and the tokenizer field is tagged only as "Other". Anyone stating the answer as a fact — in either direction — is going beyond what the data supports.

How we label sources on this page

Anonymous models attract a lot of confident numbers. We separate what TokenRA states from what a third party has measured from what is only a hypothesis.

Provider-listed stated by TokenRA or the provider, not independently verified Independent measured by a named third party, source linked Hypothesis inference, not fact

What the listing does not say

Start with the gaps, because they define how much can honestly be concluded:

FieldStatus
ProviderAnonymous by design
TokenizerClassified only as "Other"
Knowledge cutoffNot published
Parameter countNot published
ArchitectureNot published
Training dataNot published

TokenRA's own note is explicit: the model "is developed and operated by a third-party provider who has chosen to remain anonymous during this preview," and TokenRA "is not its developer, owner, or provider." The listing is the only official document, and it deliberately withholds identity.

The evidence for MiniMax Independent

The stealthprint project published a space-bunny case on 24 September 2026 using a vocabulary differential: identical probe text is sent to the mystery model and to candidate tokenizers, and token counts are compared. Query the model through its API, and the gap between candidate tokenizers becomes measurable. The results:

SignalResultWeight
Tokenizer vocabularyMiniMax 24/24Strong — and discriminating
Same run: Kimi13/24Control
Same run: GLM7/24Control
Context ceiling≥ 1M tokensRules out the m2.5 generation
Same-gateway deltaIdentity with minimax-m3 / m2.5Supports the same serving family
Wrapper overhead+143 tokens, zero driftConfirms a stable serving path
Vision stackAdapter-classReal encoder, not a placeholder
Self-identificationImprovised ChatGPT claimWeighted at zero

Why the 24/24 matters: the control numbers are what make it meaningful. If every candidate scored near-perfectly, the test would prove nothing. Kimi landing at 13/24 and GLM at 7/24 shows the differential actually discriminates — and one candidate still scored a clean sweep. Adding the 1M context ceiling narrows it further, since that excludes the earlier m2.5 generation.

Source: stealthprint case study (MIT, 24 September 2026). We did not run this test; we are reporting it with attribution and linking the raw method.

Why the self-report is worthless here

Ask this model who it is and it reportedly produces an improvised claim about being a ChatGPT model. That is not a clue — it is noise, and it is worth understanding why anyone building an identity test excludes it:

So when you see a screenshot of the model claiming to be something, treat it as a data point about the system prompt, not about the model's origin.

What would settle it

Three things, in descending order of reliability:

  1. A vendor claiming the model. This is the only definitive answer, and it is the historically normal outcome. That is documented in the stealth Alpha pattern tracked on our home page: every prior codename in this series has been claimed, usually within about a week.
  2. Reproduced fingerprinting by a second independent party. The method and tokenizer files are public, so anyone can rerun it. A second lab reaching 24/24 on the MiniMax vocabulary would harden the case considerably.
  3. Continued same-gateway comparison as MiniMax releases new named models — if the wrapper constant keeps matching, the serving-family link gets stronger over time.

What would not settle it

The base rate: this is what usually happens

Stealth releases follow a pattern. In this codename series, models appear anonymously, get tested in public for a short window, then get claimed — and the claim typically arrives within days, not months. Two recent cases:

CodenameRevealed asGap
Ox AlphaZ.ai GLM-5.3-FlashAbout 6 days; free preview ended on reveal
Union AlphaUnbiased ParetoPreview ended, model moved to a named listing

This is a base rate, not a prediction for Space Bunny Alpha. It tells you what usually happens to models in this position, not which lab is behind this one. The tracked series is longer than the two examples above; see the timeline on our home page.

Practical takeaway

If you are deciding whether to use it: the identity question is almost irrelevant to that decision. Whether this is MiniMax or another lab, the things that affect your build are measured elsewhere — the 1M context works, tool calling works, vision is real, it is free right now, and there is a ~1.8% availability gap to plan a fallback for. The strongest available usage signal is that the five heaviest consumers are all coding agents.

Identity matters for one thing: expect the free window to close when the model is claimed. That has been the pattern every time in this series. Do not build a dependency on a free preview without a fallback route. See our API guide for how to structure that.

Is Space Bunny Alpha MiniMax? FAQ

Is Space Bunny Alpha MiniMax?

Not confirmed. The best public evidence is a 24/24 tokenizer vocabulary match with MiniMax, against Kimi at 13/24 and GLM at 7/24, plus a verified context ceiling of at least 1M tokens that rules out the older m2.5 generation. No vendor has claimed the model and the provider is listed as anonymous.

Who made Space Bunny Alpha?

Unknown. TokenRA states it was developed and operated by a third-party provider that chose to remain anonymous during the preview. No company has publicly claimed it as of 24 September 2026.

Why is the tokenizer match strong evidence?

Because it cannot be spoofed by a prompt, and the probe differential cancels the chat-template constant, leaving only the model's own tokenizer count. The control scores (Kimi 13/24, GLM 7/24) show the test discriminates between candidates rather than matching everything.

Should I trust the model when it says who it is?

No. It reportedly produces an improvised claim about being a ChatGPT model when asked directly. For stealth models, self-descriptions are frequently bait, which is why serious fingerprinting toolkits deliberately exclude self-identification probes — it is the one signal that can be actively misleading while a tokenizer differential cannot be spoofed through a prompt.

Related pages