Anonymous Model Space Bunny Tops OpenRouter and OpenCode Daily Rankings, Hands-On Coding Test Finds

Discussion on X and Reddit has been mixed but leaned positive, QbitAI reported. A user posting as @CoderGeeta called the model unbelievable, while @Fei2411 said the output quality was good and @marcthecreatorr praised its speed as beating Astra. Some developers described Space Bunny as a workhorse for coding tasks; others said it spends too much time reasoning, which QbitAI's own tests also observed.

QbitAI pulled the model's API from OpenRouter and ran it inside DeepSeek Harness, a setup that has no image input, under the local nickname "tuerye." Four tasks were selected for the write-up. In the first, Space Bunny was asked to build a detailed cube-style Temple of Heaven in Three.js with interactive elements. It finished in less than two hours. Beyond the interactions requested, it added a bottom control bar with night, dawn and noon time settings plus rain and smoke weather toggles, and a roughly 3.5-minute day-night cycle whose time labels run through the twelve traditional Chinese hours. Because DeepSeek Harness cannot see images, QbitAI said it added an extra point to the model out of apology.

A second task asked for an Apple-system alarm clock. Space Bunny completed it in 22 minutes 46 seconds, producing an alarm that takes over the whole Mac screen when it rings. A follow-up request added a monthly on-off switch and moved the wake time to 7:23. The project's readme noted the model accounted for hard iOS limitations.

A third task, a Mac app that displays DeepSeek Harness's reasoning as on-screen subtitles, produced a usable first version in under five minutes, though the window could not be moved, closed or minimized. After the reviewer complained, the model fixed it in under 15 minutes while several other tasks were running. Space Bunny itself noted that synthesized mouse events are partially dropped in a headless environment, so drag synchronization with the pointer could not be asserted automatically, and the window flickers during dragging because its size changes in real time as reasoning is displayed.

In another test, the model was asked to add cursor pagination and combined tag filtering to an article list in a GitHub repository while keeping the old interface working and completing tests. It first claimed all checks passed, then reported 33 remaining lint errors; the issue was corrected once raised, and a Codex review found no problems. QbitAI also gave the model a video of a Tesla Optimus robot working a bar, sourced from @cb_doge on X, and asked what the robot was doing. Cache hit rates stayed above 95 percent on other tasks, but were unusually low on that one, QbitAI said.

Speculation about the developer was varied. Some users said the model's voice resembles GPT and guessed OpenAI; others took the word "space" in its name as a hint that it is a new Grok. Chinese labs including Kimi, GLM, HY and MiniMax were all named, and some suggested Meta's forthcoming Muse Spark.

QbitAI argued that the more consequential question than the developer's identity is why Space Bunny climbed the call rankings so quickly. It noted that most users now judge models on speed, accuracy and price, making fast Flash-class models the main tools for high-frequency work, and that Space Bunny most likely targets that segment. Its pricing will not be known until a vendor claims it.

Editor's Summary Space Bunny, an anonymous model, reached the top of the daily call rankings on OpenRouter and OpenCode within days and drew divided but largely positive developer feedback. A QbitAI hands-on test found it handles coding tasks quickly with minor first-round defects that were resolved after follow-up prompts, while its developer and pricing remain undisclosed.