On September 23, an anonymous large model named Space Bunny launched on the globally renowned model aggregation platform OpenRouter and the coding platform OpenCode.

Within just a few days, this "space rabbit," whose development team has not yet been disclosed, has climbed to the top of the daily usage charts on both platforms.

In statistics from September 27, Space Bunny's token usage on OpenRouter was approximately 3.99 trillion, and on OpenCode approximately 8.3 trillion, both ranking first.

Currently, Space Bunny has been "dominating the charts" on both platforms for three consecutive days.

According to OpenRouter, Space Bunny has a 1 million token context window, supports text, image, and video inputs, allows adjustable reasoning intensity, and provides tool-calling capabilities.

In the release announcement, Space Bunny is described as a Flash model, emphasizing fast inference and coding performance.

Developers have already used Space Bunny to complete development tasks such as web games, music players, and backend interfaces, and have given high praise after trying it out.

X creator siddharth gave Space Bunny a prompt and several images, asking it to create a seaside game called Seaside. The model quickly understood these requirements, and the generated boats, stations, and overall color scheme all met expectations. The video quickly gained over 80,000 views across the internet.

Alexey Fateev's task was to create a voxel-style pagoda garden using a single HTML file.

Space Bunny generated the pagoda, cherry blossoms, pond, and waterfall, and even added day-night switching.

After the main garden was completed in the first round, the author added another round of instructions to adjust the composition, interface, and lighting.

Although the final product had imperfections, the extreme generation speed left a particularly deep impression on him.

Beyond games, Alan tried four development tasks: writing a FastAPI interface, configuring AWS infrastructure with Terraform, creating a Stripe-style data dashboard, and setting up a ChromaDB data processing pipeline.

He said all four tasks passed on the first generation without any code modifications or additional debugging.

Developer Erick asked Space Bunny to create a Spotify-style music player.

Space Bunny quickly completed the task, and the generated result also took care of details such as fonts, album covers, playlists, current playback status, and light and dark themes.

On Reddit, some developers compared it with models such as GLM and DeepSeek.

User Eastern_Cupcake said web design was the most surprising part for them.

Several front-end cases they tried all performed quite well, with overall completion slightly better than GLM-5.3 Flash; in their own use cases, Space Bunny ran stably, followed instructions more strongly, produced fewer hallucinations, and was faster than MiMo 2.6.

User Zealousideal assigned the same code functionality review task to Space Bunny, GLM 5.3 Flash, DeepSeek 4.1, and MiMo 2.6 respectively.

They believed Space Bunny had the best understanding of context, with most of its thinking process having practical value, similar to GLM's performance.

Other developers also compared model performance in troubleshooting and fixing issues.

In a discussion thread titled "Space Bunny is better than I expected," one user said Space Bunny performed better than DeepSeek V4.1 Flash and GLM 5.3 Flash in their tests.

Another user who had done multi-model A/B testing believed it made fewer errors, was faster, and had better planning and diagnostics.

So which company does Space Bunny actually come from?

In discussions, Google, Kimi, Xiaomi, Zhipu, and MiniMax all entered the list of candidates.

Among them, MINIMAX-W (HKEX: 00100) is a name repeatedly mentioned in the community.

In a Reddit post titled "M3.1 is Space Bunny Alpha," the author described their test: sending seven different sets of strings to Space Bunny, recording the input token counts, and then comparing them with the public tokenizers of multiple model families to observe changes in token counts across different strings.

The MiniMax M series showed a variation pattern across the seven sample sets that was consistent with Space Bunny.

Independent research website YFarmX also published similar results: among its 50 test strings, Space Bunny's token counts all matched MiniMax models.

In the community, some developers also raised guesses about the model version.

Lumina said they had found traces of M3.1 in multiple file and model directory tests of MiniMax Code, which also matched the 1 million token context and Max reasoning tier configuration, all features close to Space Bunny.

Therefore, they concluded the anonymous model might be MiniMax M3.1.

Another user also mentioned in a Reddit discussion that since Space Bunny is a Flash model, the model name is very likely M3.1 Flash.

Currently, Space Bunny remains anonymous.

Behind the popularity after Space Bunny's launch is global developers' continuously rising enthusiasm for Flash models.

Domestic and international companies are intensively ramping up Flash.

Google updated its Flash series three times in six weeks, while DeepSeek V4.1 Flash, Zhipu GLM 5.3 Flash, and Xiaomi MiMo-V2.6-Flash also successively joined the competition.

Several companies' efforts are gradually converging: enhancing coding, Agent, and multimodal capabilities while placing greater importance on response speed and usage cost, enabling models to handle high-frequency, large-scale tasks.

Flash models have unique advantages for these high-frequency, everyday invocation needs.

The concentrated push into Flash models by domestic and international giants reflects that the large model industry is re-examining the development approach of purely pursuing parameter scale.

For developers who frequently use coding Agents, a model's response speed and invocation cost directly affect their daily work.

A single task may involve repeatedly reading files, modifying code, running tests, and then continuing to adjust based on errors, with each round's waiting and costs accumulating.

Whether a model is suitable for heavy use depends on whether it can complete tasks, as well as how long the whole process takes and how much it costs.

In OpenRouter's latest usage rankings, Space Bunny, DeepSeek V4.1 Flash, and GLM 5.3 Flash occupy the top three spots and form the new "big three" of Flash models. Space Bunny has successfully squeezed into this fiercely competitive market.

As for whether Space Bunny will turn out to be one of the familiar names speculated by the community or another unexpected team, it remains to be revealed.