A NEW SIGNAL FROM THE UNKNOWN
An anonymous reasoning model with a million-token context, image and video input, and a free preview. Let’s see what this bunny can do.
Independent field guide · Checked
DEMODOKOS / LOCAL AI AUDIO
Expressive voices. Original music.
Produced on your own PC.
A little ear candy.
Scene 01 · Demodokos demo
01 SPECS, SPEED & RESEARCH
Room for long documents. Speed for everyday work. Clues to the model’s origins. Explore twelve findings and the evidence behind them.
Provider specifications & telemetryIndependent testsIdentity hypothesis
Tokens are pieces of text. Context is what fits in a request. P50 means the median.
Bring a long brief, substantial documents or a sizeable codebase into one request.
OpenRouter ·
Provider-listed context capacity for stealth/space-bunny-alpha. Context capacity and completion budget are separate limits.
OpenRouter model listing OpenRouter models catalogA large output budget for long explanations and code. Reasoning uses part of that budget.
OpenRouter ·
Provider-configured maximum completion tokens. This is a budget ceiling, not a measured answer length; reasoning and visible output share the completion budget.
OpenRouter models catalog OpenRouter reasoning tokensThe median speed at which the model generates tokens. Higher throughput means less waiting for a long response to finish.
OpenRouter ·
OpenRouter P50 throughput, best across providers. Dashboard filters: all reasoning efforts and all locations. Snapshot around 00:28 UTC on 24 September 2026.
OpenRouter model listingOpenRouter’s median response delay: just over a second in this snapshot.
OpenRouter ·
OpenRouter P50 latency, best provider. Filters: all reasoning efforts and all locations. Snapshot around 00:28 UTC on 24 September 2026.
OpenRouter model listingAt least one provider returned a response for this share of requests. Errors and empty responses count against the score.
OpenRouter ·
OpenRouter’s 3-day dashboard measures whether inference returned from at least one provider. Reachability is separate at 97.70%; 24-hour inference availability is 95.60%. Snapshot around 00:28 UTC on 24 September 2026; model listed 23 September.
OpenRouter model listingThe way it counts pieces of text matched MiniMax on every test—a strong clue to the model family.
OpenCode Go ·
Published delta vectors match both named MiniMax endpoints on all 24 probes. L1 output cap: 1 token; local reference: MiniMax M1 tokenizer. This fingerprint does not establish ownership or an exact checkpoint.
stealthprint case study stealthprint measurement dataThe service usually counted 143 input tokens beyond the supplied text. This repeatable overhead helps fingerprint the route.
OpenCode Go ·
Reported prompt tokens minus the local MiniMax count of user content: +143 from the base test through the 1M-token ladder. A separate 810K retrieval probe recorded +157. This measures route-level token accounting.
stealthprint case study stealthprint measurement dataFound three different codes hidden in a very long input, and returned them in the correct order.
OpenCode Go ·
All three distinct codes were returned in order in one trial containing 200,044 local tokens and 200,187 reported prompt tokens.
stealthprint case study stealthprint measurement dataSeven text and seven image requests all succeeded, with consistent response formats and token counts within each group.
OpenCode Go ·
Seven text requests at temperature 2.0 and seven repeated 64 × 64 red-image requests. All returned HTTP 200. Prompt counts were 146 for text and 195 for images; message fields were consistent. This does not establish identical backend weights.
stealthprint case study stealthprint measurement dataCorrectly identified four red and four blue images. A simple demonstration that the endpoint handles image input.
OpenCode Go ·
Four solid-red and four solid-blue images, each 64 × 64 pixels, were identified correctly. A separate 1 × 1 red-image check returned three correct answers from four attempts.
stealthprint case study stealthprint measurement dataCorrectly named the release month of Claude Opus 4.5—a concrete example of late-2025 information it recalled.
OpenCode Go ·
Space Bunny correctly named November 2025 as the release month of Claude Opus 4.5. Anthropic’s announcement dates the launch to 24 November. This checks recall of one dated event.
stealthprint measurement data Anthropic: Introducing Claude Opus 4.5The tokenizer tests make MiniMax the family-level lead. M3.1 is a proposed identity; the developer and exact version remain unconfirmed.
Identity hypothesis ·
The study compares MiniMax M3 and M2.5 and finds a shared tokenizer signature. M3.1 is a proposed exact-version hypothesis, not a finding of that test. The developer has not revealed the model’s identity.
stealthprint case study stealthprint measurement dataIndependent probes: stealthprint on OpenCode Go’s space-bunny-free, reported 24 Sep. Provider figures: OpenRouter.
Same HTTP-routing writing task, 300-token output cap. Latest rates: Go 89.9 tok/s; Zen 74.5 tok/s. Method ↗
1.2sOpenCode Go
latest test
1.6sOpenCode Zen
latest test
02 THE SVG LAB
Pelicans, portraits and improbable passengers. See what Space Bunny draws with SVG code.
Remarkable how cute and artistic this model is drawing SVG!Generate an SVG of a pelican riding a bicycle
A raccoon sits on a dog's back while the dog steers a jet ski
Replicate mona-lisa as SVG
Create an SVG of a amazing beach scene. Sunset, animals, fun, birds, reflections
03 THE MYSTERY
The developer is still anonymous. One tokenizer clue points toward a familiar family.
Space Bunny Alpha is listed as a reasoning model from an anonymous third-party provider. OpenRouter is the gateway.
Read the listing ↗stealthprint reports 24 / 24 matching token-count probes on OpenCode Go, including comparisons with that gateway’s MiniMax M3 and M2.5.
A tokenizer-family clue. Shared tokenizers and gateway accounting can match without proving the developer or exact checkpoint.
Inspect the measurements ↗04 PUT IT TO WORK
Use the model’s inputs, tools and reasoning controls to shape your next task.
Build your first request ↓Send a screenshot, diagram or video alongside your question. Space Bunny accepts text, images and video, and replies in text.
Your application can offer functions, such as finding a file or looking up a record. The model can request the right function and supply its arguments.
OpenRouter offers low, medium, high, xhigh and max. Reasoning stays on; our code builder starts at low, while the provider currently defaults to max.
JSON response format is supported, without JSON Schema enforcement. The neural architecture, parameter count and training details remain undisclosed. Provider catalog ↗
YOUR IDEAS. A WHOLE NEW SOUND.
Turn dialogue into a performance. Add original music, expressive voices, translation and subtitles — all in one local studio.
Scene 02 · Demodokos demo
05 THE LAUNCHPAD
Choose a route, shape a request and take the code with you. Your next experiment starts here.
stealth/space-bunny-alphaPrompts and completions may be retained by the provider; they are not used for training.
Open the model ↗space-bunny-freeModel-specific documentation states zero retention and no training use. Account and route eligibility apply.
OpenCode setup ↗The example requests 4,096 output tokens. OpenRouter’s reasoning uses part of that budget.
Reasoning controls ↗curl https://openrouter.ai/api/v1/chat/completions \
-H "Authorization: Bearer $OPENROUTER_API_KEY" \
-H "Content-Type: application/json" \
-d '{
"model": "stealth/space-bunny-alpha",
"messages": [{"role": "user", "content": "Write a Python function that validates SRT timestamp order. Include three unit tests."}],
"reasoning": {"effort": "low"},
"max_tokens": 4096
}'Preview pricing checked 24 Sep 2026. Account limits and rate limits apply. OpenRouter limits ↗
06 BEFORE YOU HOP
A reasoning model from an anonymous provider, listed on OpenRouter on 23 September 2026. It accepts text, images and video and returns text, including code and tool requests. OpenCode lists Space Bunny Free under its own API model ID.
Yes, especially for creative SVG work. Space Bunny Alpha’s SVG benchmark showcases a rare artistic touch, with expressive characters, thoughtful composition and a distinctive sense of color.
The developer is unconfirmed. An independent OpenCode Go study found a MiniMax-compatible token-count signature on 24 probes. That is a family-level clue. MiniMax M3.1 is a proposed candidate, but the exact model has not been revealed.
OpenRouter currently lists $0 input and $0 output per million tokens for the preview. OpenCode describes its offer as free for a limited time. Account eligibility and rate limits still apply.
OpenRouter lists 1,000,000 context tokens and a separate maximum-completion ceiling of 524,288 tokens. Reasoning contributes to completion-token accounting. The context listing and output ceiling are capacities, not measured retrieval or generation scores.
Use Space Bunny Alpha through OpenRouter, or access Space Bunny Free through OpenCode. The API quickstart includes examples for both providers.
Space Bunny Alpha returns text, including code. Demodokos is a separate local studio for music, speech and dubbing. Listen to the demo.
OpenRouter lists reasoning as mandatory, with low, medium, high, xhigh and max efforts. The catalog default is max; our code builder explicitly starts at low. Hiding reasoning text does not disable reasoning computation.
The research cards use OpenRouter’s median dashboard figures. The expandable OpenCode comparison uses TokenDyno’s 24-hour averages from a fixed writing task. Different routes, workloads and timing definitions produce different numbers; each result keeps its source and window attached.
THE FLIGHT LOG
Model specifications, speed measurements, identity clues and four SVG experiments.