MiniMax M3.1 Flash: the "Space Bunny" steps out of the shadows

On October 19, China’s MiniMax released MiniMax M3.1 Flash, a fast, cheap model for agents, coding and computer use. The MiniMax blog also reveals that Space Bunny, the anonymous model that had been making the rounds on OpenRouter for the past few weeks, was in fact an early version of M3.1 Flash.

What’s new in MiniMax M3.1 Flash?

First, the numbers MiniMax itself gives. According to the blog post, the model:

1. Costs $0.15 per million input tokens and $0.60 per million output tokens, with cached tokens at $0.015.
2. Generates over 200 tokens per second in everyday use, with a median (P50) of over 170 tokens per second.
3. Is available right away through the MiniMax API, the M Plan subscription and the MiniMax Code app.
4. Is set to get open weights, but “later”, with no specific date or license.

The blog also announces a Pro version, coming later, which “pushes the boundaries of intelligence”.

Space Bunny?

This is the part the AI internet is talking about. Space Bunny showed up on OpenRouter as a so-called “stealth model”, meaning a model with no maker’s name attached that anyone could test. According to the MiniMax post, during that time the model processed over 200 trillion tokens.

According to Unite.AI, OpenRouter alone accounted for 56.7 trillion tokens, the most of any stealth model in the platform’s history, and Space Bunny was the number one model there for 10 days straight. It disappeared from OpenRouter on October 5 and came back on October 19, this time under its real name.

Interestingly, the OpenRouter data cited by Unite.AI shows the model clearly improved during testing: time to first token dropped from 1.95 s to 1.22 s, and throughput rose from 70 to 100 tokens per second. Anonymous testing on thousands of real users is a pretty clever way to polish a model ;)

How do the benchmarks look?

MiniMax is aiming squarely at agents. Results from the blog post:

1. DeepSWE v1.1 (coding): 73.7%.
2. OSWorld V2 (computer use): 61.7 points.
3. CyberGym (finding and confirming vulnerabilities): 82%.
4. BankerToolBench: 76.2.
5. Agents’ Last Exam: 29.6 points.

These are MiniMax’s own measurements, so let’s take them with a grain of salt. An independent ranking already exists: on the Artificial Analysis Intelligence Index the model scores 42 points. There’s a catch, though: it generated 180M tokens on that index, against a median of 99M, so it’s an unusually chatty model. Artificial Analysis also measured its speed at around 120 tokens per second, which is less than the claimed 200+.

How to try MiniMax M3.1 Flash?

The easiest way is through the MiniMax API platform or the MiniMax Code app. According to Unite.AI, OpenRouter offers the model for free for 72 hours after launch, and then at the same prices as MiniMax. The context window is 1M tokens (according to Artificial Analysis).

At these prices it’s a good candidate to plug into an agent harness, much like DeepSeek, which I wrote about in my post on free DeepSeek in OpenCode. And if MiniMax really does publish the weights, it will join the models you can run yourself, like Reflection Beam (provided, of course, that we have the hardware for it).

FAQ

Is MiniMax M3.1 Flash free?

Not permanently. Through the API it costs $0.15 per million input tokens and $0.60 per million output tokens. According to Unite.AI, on OpenRouter you can test it at no cost for the first 72 hours.

Does MiniMax M3.1 Flash have open weights?

Not yet. The MiniMax blog says the weights will be published later. No date or license has been given so far, and Artificial Analysis lists the model as closed.

What is Space Bunny?

It’s an anonymous model from OpenRouter. On October 19, MiniMax revealed that it was an early version of M3.1 Flash.