Space Bunny Alpha: Another Stealth Model, Is it Minimax?

Video authored by the channel shown, not by IANIA; all rights belong to the creator. Transcript and summary generated automatically; the original-language content is preserved.

Original language: English

Original video: Space Bunny Alpha: Another Stealth Model, Is it Minimax? · Channel: Fahd Mirza · 07:58 · Watch on YouTube

Topics: Software engineering with AI, AI models, Multimodal models, Health and science

Summary

Fahad Mirza tests Space Bunny Alpha, a free stealth model with a million-token context and native multimodal input, whose creator is unknown; he suspects it may be from MiniMax but has no confirmation. The model builds an interactive 3D suit configurator from a blank Blender scene, though the process takes a long time and browser rendering is slow. In a multilingual vision test, it appears to read an Italian newspaper headline correctly but may mistranslate it, and it acknowledges limits with low-resource languages. In a genetics test about a male calico kitten, its explanation includes some correct points but also a flawed claim; Mirza considers the response imperfect but not bad overall.

Key points

Space Bunny Alpha is available for free on various platforms and is described as having a million-token context and native multimodal input.
The model’s creator is unknown; Mirza suspects MiniMax based on matching tokenizers and error fingerprints, but says he has no confirmation.
Given a sparse prompt, the model creates a self-contained HTML 3D suit configurator using Blender.
The configurator offers suit views and customizable details such as lining, trouser cuffs, and buttons, with the displayed price changing as options are selected.
The configurator took about 50 minutes to build, and its browser rendering was slow.
In a vision-and-translation test using an Italian municipal newspaper image, the model seemed to identify the headline correctly, but Mirza suspected it mistranslated the meaning of “azzardo.”
The model admitted it lacked confidence with some low-resource languages.
In the cat genetics test, Mirza says the response had correct core genetics but made a flawed claim about a possible route to a male calico kitten.
Mirza concludes that the model tried its best and invites viewers to share their guesses about its identity.

Chapters

Start	Chapter	Summary
00:01	Introducing Space Bunny Alpha	Mirza introduces the unknown model, shares his tentative MiniMax theory, and outlines a test of its capabilities.
00:34	Building a 3D suit configurator	He asks the model to turn a blank Blender scene into a customizable browser-based suit configurator and tests its features.
04:16	Multilingual vision test	The model reads an Italian newspaper image and attempts translations into 79 languages, though Mirza questions its interpretation.
05:52	Testing genetics reasoning	Mirza evaluates the model’s explanation of how two orange cats could have a male calico kitten and points out a flaw.

Quotes

Nobody knows who built it, but I think it is MiniMax based on some of the matching tokenizers and error fingerprints, but I have no confirmation either way. — 00:01

And you see that as I select options the price changes, which is quite good. — 03:44

Because I think azzardo means gambling, not not waste hazard, but I think it got confused. — 05:20

But, other than that, I think it is not a bad response at all. It tried its best. — 07:58

Concepts and entities

Space Bunny Alpha, MiniMax, Hermes, Blender, Fahad Mirza, entrada multimodal nativa, raciocínio genético

Full transcript (original language)

00:01 There is another stealth model making the rounds across the model providers, Space Bunny Alpha. It just dropped for free on various platforms with a million token context and native multimodal input. Nobody knows who built it, but I think it is MiniMax based on some of the matching tokenizers and error fingerprints, but I have no confirmation either way. So, drop your own guess in the comments and let’s find out together by putting it through its paces. This is

00:34 Fahad Mirza and I welcome you to the channel. I also have started this free weekly AI newsletter, which you can subscribe at fahadmirza.substack.com. Let me take you straight to my Hermes agent, where I’m going to give it this prompt. If you look at this prompt, I am checking if this model can go from a blank Blender scene to a fully interactive customizable 3D suit configurator running live in browser. So, it should be using that tool and

01:05 then it should craft that single self-contained HTML file. I haven’t given given it any detail, so let’s see how it pans out. And you can see that it has built the script and now probing that Blender API plus check for hidden files. I already have in Blender installed on this system. Okay, so after 50 minutes, it says that it is done. So, let me open this in our browser. I’m just going to load it in front of

01:36 you. So, you see cutting the cloth and let’s wait for it. The first time when it runs, it is a bit slow, but I just want to show it to you how exactly it goes through the motion. On the right, we have a few of the option. On the left, it says Mason Bun. Bespoke tailoring. It is still cutting the cloth. So, I will just wait for it to finish cutting the cloth and then we will move forward. I’m showing you as is.

02:10 And now eventually you can see that the suit is loaded. I believe I can also rotate it with my mouse. And you see at the top right we have front, we have 3 by 4. I’ll just click on 3 by 4. And this is 3 by 4. Let me click on back. And by the way, it is quite slow. It moves, but it takes bit of a time because it’s a browser rendering. I have reloaded the browser just to make

02:41 sure we do it right. So, let’s click on back. So, bear with me. So, you can see that it is now doing back. Let’s do detail. So, this is a detail one. Looks pretty good. And let’s Okay, so the turntable is moving. So, this was a detail. Maybe I will just also do the shawl if I could. So, just look at here. There you go, a

03:12 hint of shawl. Could have been better, but it has just given us a shawl. Also, let me change the lining from red to maybe I’ll just go with green. There you go, it has turned green. Let’s do the white. As I said, it just takes bit of a time sometimes in my uh window, so that’s done. And then you see this is a trouser finish. I’ll just first go to 3 by 4 just to show you. This is a 3 by 4. Let me do for the

03:44 trouser maybe turned cuff. I’ll just click here. And the turned cuff is done. And you see that as I select options the price changes, which is quite good. And then of course you can change your buttons and all that stuff. The price changes and so is this image. You can also do the reset. So I just clicked on reset and it is resetting everything to the back. So look, I think it has done pretty well. Took really long time and

04:16 as I checked through the reasoning, I still suspect it is minimax, but anyway, let’s see how it goes. Let’s do another test. Next test is multilingual plus vision. So you can see that I’m handing out this model an image from an Italian municipal newspaper, a real one, and see if it can pick the headline and translate it into all these 79 different languages, real and invented, too. And this image is from a Ponti the San Felice sul Panaro

04:48 municipal newsletter, July 2026, so credit to them. Uh I just pulled it out from online and you can see that I already have started this prompt. So let’s wait for it to see how it goes. So the model was fairly quick to give the response and I think that it has correctly read the actual page headline, as you can see from this image. And if you are from Italy also, please confirm. But I think it mistranslated the whole thing. As far as I can see from Google

05:20 Translate, but please check your language, let me know. Because I think azzardo means gambling, not not waste hazard, but I think it got confused. And it is just going about it um talking about some different stuff here and there. So I’m just scrolling through. Check your language, let me know if it got it right or wrong or got totally confused. Because this is not just a mere translation task. This is um the whole context, the semantic, and

05:52 making sense out of that real world news newspaper page where there is lot happening as you can see. And yes, it corrected, you know, talked about catnip and all that. But for the low resource languages, it has admitted it has no clue. Anyway, let me know your thoughts in the comments. Let’s do another scientific reasoning test. Finally, let’s test that if this model can do real genetics. Two orange cats stranded alone for a year on an island

06:24 produced a male calico kitten. I’m asking the model, “Explain how is that possible?” And it needs to do all this reasoning in scientific lingo. I already have provided this to the model. So, let’s see how it goes about it, and we will check out if it holds up to its scientific reasoning benchmark. And by the way, if you want to help out the channel, please become a member. The response from model is actually

06:55 quite insightful and interesting. If you read through it, the core genetics are right. Two solid orange parents are locked as XY and XX. So, a normally XXY son would still be all orange, and the usual male calico Klinefelter on answer is there, which I think is not correct. I think that um what it did here is it is talking about um you know, some of the insights which I was quite impressed with that black

07:26 patches plus white spotting both require extra events not present in the island gene pool. So, this is possibly only as stacked de novo mutation, not ordinary inheritance. But I think where it slipped is that claiming a stray black cat a black parent plus sex chromosome non-disjunction can make this calico male. I told it that you know, that route is closed. And it is just a standard one, and I think that is like one in a

07:58 trillion figure which it did where it slipped. But, other than that, I think it is not a bad response at all. It tried its best. That is I must give it to the model. So, look, that is what this model is. Let me know your thoughts, especially your guesses in the comments. Thank you for all the support.