Space Bunny Alpha Review 2026: A Free 1M-Context Model That Ties MiMo-V2.6-Pro — but Two Gateways Disagree on Your Privacy
Space Bunny Alpha Review 2026: A Free 1M-Context Model That Ties MiMo-V2.6-Pro — but Two Gateways Disagree on Your Privacy
Verdict first: Space Bunny Alpha is an anonymous model that appeared on OpenRouter and OpenCode on September 23, 2026, carries a 1,000,000-token context window, accepts text, image and video, and costs $0 to call while the preview lasts. On the independent 157-test Official A board it scored 128/157 (81.5%) — an exact tie with MiMo-V2.6-Pro, and eight points behind Union Alpha despite finishing in roughly a fifth of the wall time. A 24-probe tokenizer fingerprint points at the MiniMax M3 family, though nobody has confirmed it. The headline is not the benchmark score, because there are no named competitors to score against, no published parameter count, no architecture and no license. The headline is this: the same model comes with two mutually exclusive answers about what happens to your data. OpenRouter says its Space Bunny provider may retain prompts and completions; OpenCode says its Space Bunny provider follows zero retention. Both are describing the same weights. If you would not paste something into a model whose developer you cannot name, this one has not earned that trust — and the free price is what you are being paid for the difference.
Space Bunny Alpha — a free, unnamed 1M-context model that ties MiMo-V2.6-Pro at 128/157, with a tokenizer fingerprint pointing to MiniMax
What Space Bunny Alpha Is
On September 23, 2026, a model appeared on OpenRouter and OpenCode under the identifier stealth/space-bunny-alpha. "Stealth" is a specific arrangement rather than a marketing label: a developer lists a model anonymously during a preview period, and the gateway routes requests to it without disclosing who is behind it. OpenRouter's own documentation is explicit that it is not the developer, owner or provider of the model — it is a router pointing at somebody else's endpoint.
The specification sheet is genuinely unusual for a free endpoint:
| Specification | Value |
|---|---|
| Model ID | stealth/space-bunny-alpha |
| First seen | September 23, 2026 (OpenRouter, OpenCode) |
| Context window | 1,000,000 tokens |
| Maximum output | 524,288 tokens |
| Input modalities | Text, image, video |
| Reasoning effort | low / medium (default) / high / xhigh / max — cannot be fully disabled |
| Tool calling | Yes |
| Structured output | JSON via response_format — JSON Schema not enforced |
| Pricing (preview) | $0 input / $0 output |
| Providers | ~11, including OpenRouter, OpenCode, Cline, Kilo Code, AIMLAPI |
| Parameter count | Undisclosed |
| Architecture / license | Undisclosed |
| Knowledge cutoff | Undisclosed |
Two details in that table deserve more attention than they usually get. First, a 524,288-token maximum output is four times what Ox Alpha listed in August, and it is the kind of number that matters if you are generating long artifacts rather than answering questions. Second, reasoning cannot be turned off. The ladder runs from low to max, but there is no off switch, and reasoning is always active across the full 1M context — which is part of why the model is fast and part of why its token economics are impossible to predict from the outside.
The endpoint briefly went offline mid-preview while the provider fixed an issue, then returned. That is a normal event for a stealth listing and a useful reminder of what you are actually depending on.
The Real Cost: Two Gateways, Two Privacy Answers
Space Bunny Alpha is free, which means the interesting question is what you are paying with instead. The answer differs depending on which door you walk through, and the two doors lead to the same model.
| Gateway | Stated data policy for Space Bunny |
|---|---|
| OpenCode | Provider follows zero retention and does not train on user data |
| OpenRouter | Provider may retain prompts and completions; not used for training; otherwise governed by the Stealth Model Terms |
Both statements are about the same weights. One promise is that nothing is kept; the other is that something may be. That is not a contradiction in the technical sense — different gateways can route to different serving environments, and a provider can honour different terms per contract — but it is a contradiction in the sense that matters to you, which is whether you can rely on a single answer. The practical rule: for a stealth model, the retention policy is a property of the route, not of the model. Check the specific provider you are calling, not the model name, and do not assume stealth endpoints are uniformly ephemeral because they are unnamed.
There is a second-order version of the same problem. Theneuron's analysis notes that OpenCode says data gathered during Big Pickle's free period may be used to improve the model — a different arrangement from Space Bunny's, from the same gateway. Policies vary per listing, per provider and per period. Treat any free endpoint as governed by whatever the page you are reading says today.
And a third point that tends to get lost: the free price is not a subsidy, it is a purchase. A stealth preview exists so a developer can gather real-world evaluation traffic and word of mouth before committing a name to the model. You are being paid in access for usage data and a benchmark result. That is a legitimate trade — but it is a trade, and it is worth making deliberately rather than by default.
Who Made It? The Tokenizer Fingerprint
Asking a stealth model who built it does not work. Theneuron reports that early attempts produced contradictory answers, including references to OpenAI, and notes the obvious reason: a model can confidently invent an identity. Indirect methods are more useful, and the one that produced the strongest signal here was tokenizer fingerprinting.
On September 24, an independent test pushed 24 token-count probes through OpenCode's Space Bunny route and compared the counts against other models served on the same gateway:
- MiniMax models matched across all 24 prompts — a clean sweep
- Kimi and GLM models matched less closely
That moved a MiniMax M3-family model — possibly an M3.1 or later preview — to the front of speculation, displacing early guesses that centred on Moonshot's next Kimi. The caveats are important and the analysis states them: a tokenizer match identifies a model family, not a developer or a version; tokenizers can be shared; gateways insert their own processing; and the test ran on OpenCode's route while OpenRouter lists the model under a different identifier. MiniMax has not confirmed anything.
The prior record on stealth models is genuinely mixed, which is the best argument for reading the fingerprint as a lead rather than an answer:
| Stealth model | Appeared | Turned out to be |
|---|---|---|
| Pony Alpha | February 2026 | An early GLM-5 test |
| Hunter Alpha | March 2026 | Xiaomi MiMo-V2-Pro — the crowd had bet on DeepSeek and was wrong |
| Ox Alpha | August 20, 2026 | Z.ai / Zhipu GLM-5.3-Flash, confirmed August 26 |
| Union Alpha | September 16, 2026 | unbiased.ai "Pareto 26.9", identified in about a day |
| Space Bunny Alpha | September 23, 2026 | Unconfirmed — MiniMax M3 family per the fingerprint |
The Union Alpha case is the instructive one, because it shows how quickly and how cleanly a stealth model can be unmasked when there is a matching artifact to find. A researcher matched Union Alpha's completion-ID pattern against The Unbiased Co's own platform and observed requests being answered by the same vLLM fleet. Within roughly a day the stealth listing was replaced by a paid unbiased/pareto entry. Note also how Union Alpha went out: it consumed 2 billion tokens on its first day, with early testers claiming frontier-adjacent performance. Stealth previews are now a recognised launch strategy, not an oddity — which is exactly why the crowd sometimes nails the attribution and sometimes hands the wrong company free publicity.
For Space Bunny Alpha specifically, the fingerprint is corroborated by circumstantial fit. MiniMax's M3, released June 1, 2026, is a 428B-parameter MoE with roughly 23B active, a 1,048,576-token context, native text-image-video input, and a sparse-attention architecture built for long context. Space Bunny's advertised 1M context and video input line up with that profile. MiniMax is also reported to be building a 2.7-trillion-parameter model, codenamed M3 Pro, targeted for open-source release as early as Q3 2026. A stealth preview in late September would sit neatly in that window. None of this is proof, and it should be read as a coherent hypothesis rather than a finding.
The Benchmarks: 81.5%, and the Shape of the Failures
Because there is no Artificial Analysis listing and no arena rating for Space Bunny Alpha, an independent third-party board is the only real evidence available. The Official A board ran 157 tests across six categories with thinking disabled, max_tokens capped at 1,024, temperature 0 on the smoke test, and a text-only path.
| Model | Score | Fails / errors | Wall time |
|---|---|---|---|
| Union Alpha | 136/157 (86.6%) | 21 / 0 | 16,540.5 s (275.7 min) |
| Space Bunny Alpha | 128/157 (81.5%) | 29 / 0 | 2,896.6 s (48.3 min) |
| MiMo-V2.6-Pro | 128/157 (81.5%) | 29 / 0 | 2,536.6 s |
| Ox Alpha | 129/157 (best thinking-off) | — | Not like-for-like |
The MiMo-V2.6-Pro tie is the useful anchor because that model is one I have covered before: a 1.02-trillion-parameter MiMo flagship, trained in about six days for a reported $3.47M, that sits at the top of the open-weights rankings. A free, unnamed model matching it on aggregate is a real result. The Ox Alpha row is deliberately not a like-for-like comparison — that listing used a different 4,096-token path and sits in a different reasoning category — and the analysis declines to publish a single overall rank, which is the right call.
The aggregate hides the more interesting story, which is in the per-category breakdown:
| Category | Space Bunny Alpha | Union Alpha | MiMo-V2.6-Pro |
|---|---|---|---|
| Instruction following | 28/30 | 28/30 | 28/30 |
| Prose | 23/30 | 24/30 | 28/30 |
| Reasoning | 24/30 | 30/30 | 24/30 |
| Coding | 23/30 | 26/30 | 25/30 |
| Math | 23/30 | 24/30 | 18/30 |
| Writing | 5/5 | 2/5 | 3/5 |
| Tool calling | 2/2 | 2/2 | 2/2 |
Three things stand out. Space Bunny sweeps the small writing suite (5/5) where Union Alpha manages only 2/5 — a reminder that a single aggregate number can invert on a subtest. It beats MiMo-V2.6-Pro on math (23 vs 18) while tying it overall. And it is six points behind Union Alpha on reasoning, the largest single gap in the table, which is where Union Alpha's eight-point overall lead actually comes from.
The failure shape is the most actionable finding here. Of the 29 failed tests, the breakdown is unusually diagnostic: 7 coding failures were overwhelmingly syntax errors — five invalid decimal literals and one unterminated string, plus one undefined name; 7 math failures were exact-decimal format mismatches; 6 reasoning failures were boxed-answer or integer format mismatches; and the instruction and prose failures came from exceeding a requested line count — one test asked for 6 lines and received 129. That is not a model that cannot reason. It is a model that struggles with output formatting conventions, particularly numeric literals and exact-format answers. If you are parsing its output programmatically, budget for validation layers around numbers and strict formats — and note that the one-instance-asked-for-6-got-129 error is the kind that breaks a downstream pipeline rather than just disappointing a reader.
On latency, the timing data is more encouraging than the raw wall time suggests: mean 18.43 s, median 11.59 s, maximum 85.97 s, with seven tests past 60 seconds and none approaching the 300-second timeout. The run was explicitly not a timeout run. The honest summary is that Space Bunny Alpha is roughly 5.7× faster than Union Alpha on this board for eight points less accuracy — a speed-accuracy trade that will look very different depending on whether you are running an interactive tool or a batch pipeline.
Free Now, Priced Later
The obvious question about a free frontier-adjacent model is how long it stays free, and the available precedent is not encouraging for anyone planning to build on it. Big Pickle, another free stealth model listed by OpenCode, has been running since October 17, 2025 — more than eleven months — and is still labelled a free, limited-time stealth model, with a 200K context and a 32K output cap. Its identity was never confirmed either; rumours point at a Zhipu GLM-4.6 variant with others suggesting DeepSeek for later versions.
Two lessons follow. First, "limited-time" is not a schedule — it can mean a week or a year, and the developer sets it unilaterally. Second, an unresolved stealth model can persist indefinitely, which sounds like good news for free access and is actually bad news for anybody who needs to plan capacity, negotiate a contract or make a compliance argument. Union Alpha's path is the other template: free stealth preview, then within a day a paid listing under a real name. Both endings are available and you do not get to choose which one you get.
What you cannot do with Space Bunny Alpha today is any of the things that make a model a component rather than a curiosity. There is no parameter count, no architecture, no tokenizer, no knowledge cutoff, no license, no service-level commitment and no stated deprecation policy. Those absences are not unusual for a stealth preview, but they are the reason it belongs in an evaluation harness and not in a production dependency graph.
The Honest Take
Space Bunny Alpha is the most interesting free model on the market right now, and the fact that it is interesting is a warning rather than a recommendation. The capability is genuine: 128/157 ties a trillion-parameter open-weights flagship, the 1M context and 524K output are real, the video input is real, and it is free. If you want to know what the current capability frontier feels like without spending anything, this is the cheapest way to find out, and the speed makes it pleasant to use.
Everything past that is unresolved. You do not know who built it. The best available evidence says MiniMax, and that evidence is a tokenizer fingerprint with three acknowledged caveats attached. You do not know what it costs to run in the sense that matters, because there is no rate card, no token accounting you control and no way to forecast whether it will exist next month. And you do not get one consistent answer about your data: the same model is described as zero-retention at one gateway and possibly-retaining at another.
Use it if: you are evaluating capability, prototyping something you were going to throw away anyway, running batch work where outputs are verified downstream, or benchmarking against your own stack — the free price and the 1M context make it an unusually good deal for exactly these jobs.
Don't use it if: you would be sending customer data, proprietary code or anything regulated — the retention answer depends on which gateway you picked and OpenRouter's own terms say prompts and completions may be kept; if you need reproducible behaviour, since an unnamed endpoint can be swapped, repriced or withdrawn without notice and the model itself has no version you can pin; or if you are parsing numbers and strict formats programmatically — five of its seven coding failures were invalid decimal literals, and that is a specific, repeatable weakness rather than noise.
Watch two things. First, whether anybody confirms the identity — the Union Alpha precedent says an attribution can land within a day once someone finds the right artifact, and a confirmed MiniMax M3-family model would immediately give this a real price list, a real license and a real knowledge cutoff. Second, whether it survives its own preview. Union Alpha became a paid listing within a day; Big Pickle has been free for eleven months. Until one of those happens, the correct posture is to enjoy the free access and build nothing that depends on it.
FAQ
What is Space Bunny Alpha?
An anonymous large language model that appeared on OpenRouter and OpenCode on September 23, 2026 under the identifier stealth/space-bunny-alpha. It is a stealth release — the developer stays unnamed during the preview, so OpenRouter routes requests but is not the model's developer or owner. It offers a 1,000,000-token context, 524,288-token maximum output, text, image and video input, tool calling, structured output and a five-level reasoning-effort control. It is free at $0/$0 while the preview lasts.
Who made Space Bunny Alpha?
Unconfirmed, and the strongest evidence points to MiniMax. A September 24 tokenizer fingerprinting test pushed 24 token-count probes through OpenCode's Space Bunny route; the counts matched MiniMax models across all 24 prompts, while Kimi and GLM matched less closely — pointing at the MiniMax M3 family, possibly an M3.1 or later preview. Caveats: a tokenizer match identifies a family rather than a lab or version, gateways can insert their own processing, and MiniMax has not confirmed. Early guesses centred on Moonshot's next Kimi.
Is Space Bunny Alpha actually free?
Free to call right now at $0 input and $0 output across roughly eleven providers. Whether it stays free has an unhappy precedent: Big Pickle, another free OpenCode stealth listing, has been running since October 17, 2025 — over eleven months — still labelled limited-time. A stealth preview exists so the developer can gather evaluation data before committing a name, and both the access and the price can change once it is named.
How good is Space Bunny Alpha on benchmarks?
On the independent 157-test Official A board, run with thinking off and a 1,024-token cap, it scored 128/157 (81.5%) with zero errors — an exact tie with MiMo-V2.6-Pro. It trails Union Alpha at 136/157 (86.6%) by eight points, but Union Alpha took 275.7 minutes against Space Bunny's 48.3 minutes, roughly 5.7× longer. There is no Artificial Analysis listing and no arena rating, so these figures stand alone rather than in a cross-lab comparison.
What are its weaknesses?
The failure profile is specific rather than general. Of 29 failed tests: 7 coding failures were mostly syntax errors — five invalid decimal literals and one unterminated string, plus an undefined name; 7 math failures were exact-decimal format mismatches; 6 reasoning failures were boxed-answer or integer format mismatches; and instruction/prose failures came from exceeding a requested line count — one test asked for 6 lines and got 129. That points to formatting and numeric-output discipline rather than reasoning depth.
Is it safe to send data to Space Bunny Alpha?
It depends on the gateway, which is the most important thing to understand about this release. OpenCode states its Space Bunny provider follows zero retention and does not train on user data. OpenRouter states its Space Bunny provider may retain prompts and completions, though not for training. Same model, two different answers. Verify the specific provider's policy rather than assuming stealth endpoints are uniformly ephemeral.
Should I use it for production work?
It is reasonable for experimentation, prototyping, evaluation runs and high-volume tasks where output is verified downstream. It is a poor production dependency: the developer is unknown, the parameter count, architecture, tokenizer, knowledge cutoff and license are all undisclosed, there is no service-level commitment behind a free stealth endpoint, and access can be withdrawn or repriced without notice. Treat it as a way to test the current capability frontier at zero cost, not as infrastructure.
Sources
- The Neuron — Who made Space Bunny? A MiniMax clue sharpens the mystery
- SMF Clearinghouse — Official A: Space Bunny Alpha 128/157, tied with MiMo-V2.6-Pro
- SMF Clearinghouse — Official A: Union Alpha scores 136/157 on a free stealth endpoint
- OpenRouter — Space Bunny Alpha: API pricing and providers
- AI/ML API — Space Bunny Alpha documentation
- GIGAZINE — The mysterious high-performance AI model Union Alpha is released, benchmark scores on par with GPT-6 Astra and Claude Fable 5.1
- 36Kr — Mysterious Union Alpha surprise launch: 2B tokens consumed on day one
- Daily AI World — Union Alpha is Pareto 26.9: Astra-level scores at 1B tokens per minute
- YFarmX — Who identified Union Alpha and Ox Alpha first: every measurement and timestamp
- OrcaRouter — What is MiniMax M3? MiniMax's 1M-context flagship model
- The Next Web — MiniMax plans China's biggest AI model, and will open-source it
- MiniMax — Official model release notes
- LLM24 — Space Bunny Alpha stealth model price and provider availability
Space Bunny Alpha's developer has not been confirmed. The MiniMax attribution rests on a third-party tokenizer fingerprinting test and should be read as a hypothesis, not a finding. All benchmark figures come from the independent Official A board and were run with reasoning disabled and a 1,024-token cap on a text-only path; they have not been reproduced by Artificial Analysis or any arena, and none exists for this model. Data-retention terms quoted here reflect what OpenCode and OpenRouter stated at publication and are governed by their respective Stealth Model Terms, which can change. Free stealth access, provider lists and pricing can be altered or withdrawn without notice — verify against the provider's own page before relying on any of it.
评论
发表评论