An unnamed model showed up on two AI platforms on September 23. It was called Space Bunny.
It was free. It could read a million characters at a time. It could look at images and videos, and you could adjust whether you wanted quick answers or more deliberate thinking.
It became the most-used model on both platforms. Nobody knew who made it.
But the model had one telltale habit it couldn’t change. Models from different companies tokenize text differently. Feed them the same sentence, and one might split it into 20 chunks, while another splits it into 23—and at different points. A model can change its name, but not the way it tokenizes text.
Developers tested it and found it matched MiniMax’s M series. MiniMax said nothing.
At the time, they also had a public beta running: M3.1-Flash-Preview. It had a name and documentation, and you had to join a queue to use it.
The two models ran side by side for a month. One had a name. Nobody even knew who made the other.
At 5:21 p.m. today, MiniMax posted: “Welcome back, Space Bunny. M3.1-Flash is coming.”
It has a name now. But no price. No launch date. And no word on whether it’ll be open source.
The pricing page is blank. People say it’ll cost $0.165 per million characters of input and $0.66 per million characters of output, but the company hasn’t confirmed a thing. On the other platform, it’s still free.
That month of free access brought back more than just users. A free model attracts bargain hunters, people curious to try something new, and people with too much work to afford paid tools. As they use it, they generate a lot of real-world data: how much they write, whether they send images, whether they want speed or accuracy. That data will shape future pricing.
The price isn’t missing because they haven’t figured it out. They just hadn’t finished doing the math. Now they have.
The people who came for free weren’t there because they loved the model. How many will stick around once it starts charging is another matter.
At the same time, a bunch of small paid models were also trying to attract users. They didn’t know who they were up against—the free, useful, nameless rabbit was siphoning away all the usage, and there wasn’t even anyone to complain to.
So the comments under this news will probably split into two camps. One will call it clever: starting from nothing and pulling off a big splash. The other will say it’s a little shady: winning by staying hidden.
I lean toward the first camp. But if you’ve used it, you’ll probably find the second more convincing.
It was free. It could read a million characters at a time. It could look at images and videos, and you could adjust whether you wanted quick answers or more deliberate thinking.
It became the most-used model on both platforms. Nobody knew who made it.
But the model had one telltale habit it couldn’t change. Models from different companies tokenize text differently. Feed them the same sentence, and one might split it into 20 chunks, while another splits it into 23—and at different points. A model can change its name, but not the way it tokenizes text.
Developers tested it and found it matched MiniMax’s M series. MiniMax said nothing.
At the time, they also had a public beta running: M3.1-Flash-Preview. It had a name and documentation, and you had to join a queue to use it.
The two models ran side by side for a month. One had a name. Nobody even knew who made the other.
At 5:21 p.m. today, MiniMax posted: “Welcome back, Space Bunny. M3.1-Flash is coming.”
It has a name now. But no price. No launch date. And no word on whether it’ll be open source.
The pricing page is blank. People say it’ll cost $0.165 per million characters of input and $0.66 per million characters of output, but the company hasn’t confirmed a thing. On the other platform, it’s still free.
That month of free access brought back more than just users. A free model attracts bargain hunters, people curious to try something new, and people with too much work to afford paid tools. As they use it, they generate a lot of real-world data: how much they write, whether they send images, whether they want speed or accuracy. That data will shape future pricing.
The price isn’t missing because they haven’t figured it out. They just hadn’t finished doing the math. Now they have.
The people who came for free weren’t there because they loved the model. How many will stick around once it starts charging is another matter.
At the same time, a bunch of small paid models were also trying to attract users. They didn’t know who they were up against—the free, useful, nameless rabbit was siphoning away all the usage, and there wasn’t even anyone to complain to.
So the comments under this news will probably split into two camps. One will call it clever: starting from nothing and pulling off a big splash. The other will say it’s a little shady: winning by staying hidden.
I lean toward the first camp. But if you’ve used it, you’ll probably find the second more convincing.