Space Bunny Alpha › What Is It?
What Is Space Bunny Alpha? Facts and Identity Evidence
Space Bunny Alpha is a free, anonymous large language model offered through TokenRA's stealth programme, released on 23 September 2026. It carries a 1,000,000-token context window, accepts text, images and video, supports tool calling, and has no disclosed owner — no model card, no paper, no published benchmark table. This page lays out what the listing states, what third parties have measured, and what the model claims about itself, with the source of each layer labelled.
Last updated 25 September 2026 · Released 23 September 2026 · Model ID space-bunny-alpha
How we label sources on this page
An anonymous model produces three very different kinds of information. Keeping them apart is the whole point.
What Is Space Bunny Alpha as a Stealth Model?
A stealth listing is a public test run under a codename. A lab puts a model into a shared gateway, the provider name is withheld, the model is offered free or cheap, and users hammer it for a few days. If the model performs well, the lab attaches its name to the result; if not, the codename is quietly retired. TokenRA describes the arrangement in its own terms: 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."
Two practical consequences follow, and they matter more than any spec. First, there is no authoritative document — no model card, no technical report, no evaluation appendix. Second, the weights are not yours: this is an API listing, not an open-weight release, so nothing about it can be inspected directly.
The listing facts Provider-listed
Everything below is stated on the model's TokenRA listing as of 24 September 2026.
| Property | Value |
|---|---|
| Model ID | space-bunny-alpha |
| Released | 23 September 2026 |
| Developer | Anonymous third-party provider |
| Price | Free (no charge for prompt or completion tokens) |
| Context window | 1,000,000 tokens |
| Max output | 524,288 tokens |
| Input | Text, images, video |
| Output | Text |
| Tool calling | Yes — accepts tools and tool_choice |
| Structured output | response_format JSON supported, without JSON-schema enforcement |
| Reasoning | Adjustable reasoning effort |
| Provider / gateway | Anonymous third-party provider; TokenRA is a routing gateway. Route topology and any fallback are not disclosed. |
The provider is anonymous and the TokenRA page does not disclose whether the route has fallback providers. OpenRouter publishes its own gateway metrics and fallback behavior, but that separate listing does not establish the TokenRA route topology.
Live service metrics Listing-reported
These operational figures are displayed by the OpenRouter listing and were observed on 25 September 2026. They are separate from the TokenRA-listed model specifications, are not quality scores, and are not independent measurements by this site.
| Metric | Value |
|---|---|
| Throughput | 80 tok/s (P50, best across providers) |
| Latency | 1.05 s (P50, best provider) |
| Uptime (3 days) | 99.76% |
| Availability (3 days) | 98.24% |
| OpenRouter availability (24 hours) | 98.57% |
Uptime and availability are different measurements from the OpenRouter gateway listing. It defines uptime as the percentage of the past three days when at least one provider responded, while availability is the percentage of time inference was successfully served, and says it can use the next-best provider when one returns an error. In this snapshot, the values are 99.76% uptime, 98.24% three-day availability and 98.57% OpenRouter availability over the last 24 hours. These figures do not describe TokenRA or prove that the two routes share an upstream provider.
Who is actually using it Provider-listed
The listing publishes its heaviest traffic sources, which is an unusually direct signal of what the model is being used for. As of 24 September 2026:
| App | Traffic | What it is |
|---|---|---|
| Cline | 54.1B tokens | Open-source coding agent that runs inside the IDE |
| Claude Code | 50.4B tokens | Anthropic's agentic coding tool |
| Hermes Agent | 32.1B tokens | Nous Research agent with persistent memory and 40+ tools |
| DeepSeek Harness | 30B tokens | Newly listed in this window |
| Kilo Code | 28.9B tokens | Open-source coding agent for VS Code, JetBrains and CLI |
The composition is the story: the top five senders are all coding agents. Whatever the model actually is, its real workload is long, tool-heavy repository work — not chat, and not one-shot generation. That is also the workload that explains the 1M-token context and the 524,288-token output ceiling.
What is deliberately not disclosed
The absence list is as informative as the specification list:
- Developer and owner — withheld by design during the preview.
- Tokenizer identity — the listing does not name it.
- Knowledge cutoff — not published.
- Parameter count, architecture, training data — none of it appears anywhere.
- Model card, paper, evaluation table — does not exist. Any MMLU, SWE-bench or AIME score you see quoted for this model is inherited from something else or invented.
What third parties have measured Independent
One public project has published a structured fingerprint analysis of a same-codename model: stealthprint, an open toolkit (MIT) for analysing stealth models, with the raw measurements published alongside the write-up. Its findings, dated 24 September 2026:
| Finding | Result |
|---|---|
| Tokenizer family | MiniMax vocabulary — 24/24 probes exact, MAE 0.00 |
| Family comparison | Kimi 13/24, GLM 7/24, Llama 4 13/24 — all worse fits |
| Chat template overhead | Fixed +143 tokens, no drift from 10¹ to 10⁶ tokens |
| Context | ≥ 1M tokens confirmed by ladder (100K / 204.6K / 450K / 1M all returned 200) |
| Long-context retrieval | 3/3 distinct needles at 200K depth, character-exact |
| Vision | Flat +34 token overhead regardless of resolution; 64×64 colour test 8/8 |
| Backend pool | 14/14 identical payloads returned identical output shape — homogeneous |
| Effective knowledge | ≥ November 2025 (recognises Claude Opus 4.5's release month) |
The limit of this evidence — read before quoting it
This study was performed on a same-codename model served through OpenCode Go, not on this TokenRA listing. The write-up states explicitly that the tested model was not listed on OpenRouter. Treat the findings as family-level evidence about a codename, not as direct measurements of space-bunny-alpha: the two may be the same model, a sibling checkpoint, or two different models sharing a name. The tokenizer result in particular is family-level by the project's own note — MiniMax's vocabulary is shared across generations, so it identifies a family, never a specific checkpoint.
What the model says about itself Hypothesis
Asked who it is, the same-codename model improvises "I'm ChatGPT, made by OpenAI." That is not a disclosure; it is what a model says when its chat template supplies no identity and it has to fill the gap. The fingerprinting write-up notes the model's own reasoning leaks the mechanism — a line about no model identity having been provided.
This is why self-report is worthless for identifying a stealth model, and why every claim on this site is labelled by how it was obtained rather than by how confident it sounds.
What it cannot tell you
- Who made it. Nothing public establishes this. The tokenizer evidence narrows it to a family, not a vendor, and not a checkpoint.
- How good it is. No evaluation table exists. Operational speed is not quality.
- How long it stays free. Free stealth previews have ended when a vendor claimed the model — the previous one, Ox Alpha, was free for roughly six days before Z.ai revealed it as GLM-5.3-Flash; Union Alpha was claimed as Pareto by Unbiased and its free period has now ended.
- Whether it is suitable for production. With a ~1.8% availability gap in the cited OpenRouter snapshot, undisclosed TokenRA route topology, and prompts retained by an anonymous operator, the honest answer for critical workloads is that the risk is unquantified.
Frequently asked questions
What is Space Bunny Alpha in one sentence?
A free, anonymous large language model on TokenRA with a 1M-token context window, image and video input, tool calling, and no disclosed owner.
Who made Space Bunny Alpha?
Unknown. The best public evidence points to the MiniMax tokenizer family, but the stealthprint case study tested a same-codename model served through OpenCode Go, not directly on the TokenRA listing. The case study says the tested model was not listed on OpenRouter, and no vendor has claimed this one.
Is it free?
Yes, as of 24 September 2026 — the listing price is zero. Free stealth previews have historically ended when a vendor claimed the model.
What is the context window?
1,000,000 tokens, with up to 524,288 tokens of output.
Does it support tool calling?
Yes — tools and tool_choice for function calling, plus response_format for JSON output without schema enforcement.
What can it take as input?
Text, images and video. Output is text.