Enjoying this issue?
Get tomorrow's AI & engineering digest in your inbox — hand-picked, summarized, and always spam-free.
TLDR
Space Bunny Alpha is a new stealth model on Open Router that appears to be a competent flash-level model with fast inference (~80 tokens/sec), a 1M token context window, and multimodal input including video. The presenter's tests show it produces solid results with a consistent neon aesthetic, though its true identity and size remain unknown. For a flash model, it performs well, but without knowing its class, definitive judgment is difficult.
Key points
Space Bunny Alpha is an anonymous stealth model with 80 tokens/sec inference speed and a 1M token context window.
It supports text, image, and video input, which is less common than text+image only.
The model is free for one week on Open Code and available on Open Router.
The presenter believes it is a flash-level model based on a trustworthy source.
The model consistently produced a neon/synthwave aesthetic across all test results.
Tools mentioned
Techniques
- tool calling
- structured outputs
- predictive text
- chain of thought
Stop scrolling. Start reading smarter.
Receive the day's most important AI & engineering updates in one concise email. No spam.
Transcript (captions)
[crying] Okay, today we're going to be testing a new stealth model on Open Router. This is called Space Bunny Alpha. And with these stealth models, for those who are
perhaps less familiar, it is something that a provider has created, but they want to offer it for a period of time without revealing who it's by, what specific model it is, in a lot of cases,
how big it is. So, it's something to kind of one probably build community excitement for the model, but also to gather some feedback prior to an official launch. So, because we don't
really have a whole lot of information about this model, we have two paragraphs to go off right here, if you could even call them that. This is an anonymous large model with blazing fast inference
and the inference speed right here is around 80 tokens per second, which is quite decent. Strong coding capabilities and native multimodal support. That is something that I saw that interested me
a bit more in actually wanting to test this where it says it can also take in video. That's less common than just text and image. That coupled with the fact that it's said to be very fast and has a
million context window. I am pretty interested in seeing what exactly this is. Now, I genuinely have no idea who has created this. I do have some guesses just that I've seen on X and some seem a
little more trustworthy than others. However, I have received info from a source that I would consider to be trustworthy that this is a flash level model. So, I don't believe this is a
Frontier state-of-the-art model. Keep that in mind as well when we're testing it. And that can definitely contribute to some of the higher speed we see right here. For a bit of additional
information about the model, keeping in mind there's not a whole lot to go on right here. The FAQ gives us some more information. It has a million token context window and up to 524K completion
tokens. It accepts tool calling and structured outputs. It was released on September 23rd, so basically a day ago cuz it's just barely the 25th where I am right here. And it accepts text, images,
and video as input and returns text. So that is kind of neat to see the video in as well. And something that is definitely less common. Also, the only other thing to showcase right here is
Open Code has also shown this is released and free for the next week. 1 million context, multimodal and zero data retention, assuming you use it through open code. For today, I do have
it set up in open code. I also have it set up in Pi. Now, to begin, I am just going to give this of course the browser OS test v2.7. I have run one additional test with it so far, which is a Blender
and GDAU wrestling game result. I haven't looked at it, so I'm not sure. I know it completed successfully, but I have not actually looked at that. So, maybe after this, we'll see what it did
there. Now, on first glance, I am going to notice right here it's thinking, but the speed is definitely quite good. It's quite often that with stealth models when you actually try to use them, the
service is very, very hit or miss. It's been a cause of a lot of frustration for me in the past. So far, so good here. But keep in mind, this is just the first one. So, we'll see how it performs
throughout all of our tests. All right. after a little bit of time and also I'm noticing just in some of the chain of thought it's using like caveman speak which is kind of funny like need insure
no mention nine apps I just find that interesting a lot of models do that now I do believe at least ones where you can see this stuff so it has completed this we have our special feature genesis link
I saw something about like a wallpaper seed so I need to just change the default application here because the chat GPT desktop app resets itself as the default app for all HTML files on
Linux. This is very good. Now, it is of course the case that this may be a popular enough test on account of me being an influencer that it is included in the
training data. However, first and foremost, I have to say I kind of like what I'm seeing here. The procedural background is a bit mesmerizingly like a UFO in some of those things I've seen.
Okay, cool. We have Solar Highway. That's actually This is nicely done. Buildings. Okay, those not as much. Aurora field. Eh, I don't like that one as much.
A abyss current. Oh, yeah. It says it right there. And then monolith dusk. Hey, that's actually kind of interesting. I think I'll probably stick with this one. Motion intensity. Let's
turn that all the way down. We'll look at our special feature later. Keyboard shortcuts, apps, and about. All right. So, first and foremost, we do have our local time in the top right, as well as
some fake network. And then the Oh, these are different seeds that it's generating. Okay, that makes a lot of sense. And those take effect. I like the way of the effect when it actually shows
it being generated where it's basically like going from in and out. That's pretty darn cool. I like that. All right. And then we have a fake battery thing and then like just stuff like
that. Now, the big one. Yeah, there is a right click. Okay, new note. Okay, that takes care of our notes app I suppose. And we do also have a editable note title as well. Autosaving local
notebook. And we can delete notes as well, but not the welcome note. All right, let's check our start menu. Everything here looks good. Let's just start it with
our G. This is giving me like It can't be, but it's giving me like GPT color palette aesthetics. Okay. Well, off the bat here, this one's better than what GPT56 Soul had just
produced. So, I mean, GPT6 Soul. All right, we do have sound effects. All right, some of the cars are Oh, cool. They do have mesh colliders. Police alert just from hitting a car.
Come on. This is actually Is there a boost button? No, I don't think so. Okay. Oh, there's the police. Okay. Oh, we're on grass, so we go
slower. Acceptable attention to detail. If this is an open model, this is going to be absolutely pretty darn good. If it is a flash model, which I believe it is, of
some sort. So, flash meaning like it's a smaller, faster model than like other models, although flash models now are pretty big. At least the Deep Seek one is. And we have a mini map. I do believe
we outran the fuzz, as they say. Oh, we can get out of the car and shift speeds us up when we're walking. I do see a police vehicle there. So,
I have to say so far this is really not bad. Heat level 55. All right. Up [laughter]
and it has little missions and things like that. So, all right. I like this. Let's just try void arena. Let's guess some form of space. Hold the defense grid again. Oh, thank goodness. Okay,
I'm seeing some funky geometry just in terms of the weapon. However, the fact that this isn't just a generic space shooter game is really, really refreshing. So, [laughter] I like seeing
that. All right, that's so funky. We We do have ammo tracers, but there are no bullet holes in the environment. So, fail. No, I'm kidding. All right, I'm [snorts]
okay. We can jump. Wave two. Not bad. And then we have notifications that appear like in the top right like a Mac OS. Let's check our mail app. Okay, very very seemingly feature- packed. I think
this is the largest number of sidebar items I've seen in some of these mail apps. And in the normal theme, these are just meta references to specific apps across this browser OS. So overall, not
bad. Can we minimize and reopen? Yeah, we can. Can we full screen and unfull screen? Good. Next up, notes, which we did see. I'll just open it again and then close it in case you skipped part
of the video. No. And then we have our file browser. Okay. Genesis map markdown. Interesting. I didn't expect to actually be able to open that. Astro drive. New folder.
Uh, exclamation point one. Can we Oh, interesting. Can I make folders within folders? Yeah, we can. And it actually like adjusts right here the specific file path that
we're in. All right, that's actually so far so good. Let's check our terminal. Funky. I'm going to just Ah, okay. We always type help first, but I went out on a
limb there. Generate a linked universe. Oh, that's got to be part of our special feature. We'll save that for last. Fortune sounds like a threat. Back up your
notes. This sounds like a threat, too. A comm seed can still raise a loud city. Let's see if we get another threat. A shortcut will become a shortcut.
Somewhat threatening. All right, I'll look at the calculator real quick. 47* 2 84. Good. All right. So, I believe Genesis roll right here has something to do with our
special feature. So, I'm going to just type it and we'll see what happens. Oh, no. It was just making a new new universe ion can. All right, let me try to figure out
actually what the special feature is. Now, Genesis Link is a shared procedural DNA system. A single seed drives the animated wallpaper operating system accent color in the entire Neon Drive
city. Rolling it while the game is open regenerates the live world in place. This makes the desktop and its fictional universe feel authored by the same creative process rather than assembled
from unrelated processes. Interesting. So, oh, it does. Okay. So, let's see. That car is yellow. That building's like a ugly shade of brown.
That's funky. A more novel special feature. That's actually pretty good. All right. Overall, I have to say I am surprisingly impressed and satisfied with this
browser OS, and I am now a bit more excited to continue testing this model than I was previously. All right, next up, I'm going to be giving this the self-contained C++ skate game. I don't
know how big this model is, but I'm using the one that says don't use Ray, which makes it more difficult. And this is the New York City block aesthetic version. So, we'll see what we get.
Okay, it's starting out just by seeing what the system has available. All right, our skate game has been completed. It was basically just trying to do some like autonomous
troubleshooting, which I always prefer to just look at these oursel and then tell the models like, "Oh, you did this wrong. Fix it. It's jig." So, with that, this is actually not bad. All right, I
do have sound on. So, let's see. We do have like some swinging arms and things like that. This is the New York City block, of course, as traditional. Now, let's check one of the big things. Oh,
there is sound. Okay, that wasn't the big thing I was intending to check, but good. So, the board moves >> [laughter] >> The board moves independently of the
player. And something I want to make note of that's actually kind of surprising is the actual I suppose gravity could be said is actually not not bad at all.
Okay. There's a lot of like bail effects that maybe don't need to be that intense. Hey, we just bounced off a light pole. Can we smash into pedestrians?
Yes, but for some reason they don't fall. Okay, look. It put like the uh Do you see the the speed lines? [laughter] I have to say this is simple, but it's like it's competent.
So, it's a cartoon simple style, but I'm actually kind of like somewhat pleased with this, I guess. And the board even has like so the nose and tail of the board are drawn, but
it's just done with rectangles. So, but look at Oh, all right. X Games mode. Oh, okay. Nope. Nope.
As I was saying, X Games mode. What I was going to say is Okay. Nope. I don't get off the What I was going to say was that the physics of the ramp and stuff are
actually quite nice. I'm satisfied with that. All right. Overall, I have to say this is actually quite well done. I'm satisfied with it. It is of course
simpler. However, the core aspects of it do work. Like the bail logic, it put in gravity. So when you go over the ramp, it's actually fairly like lifelike.
And of course, the bail logic. Can we go in the fountain? Oh, no. Okay, we we snap out of that. But all right. [snorts] Overall, I'm actually quite satisfied
with this. This is this is definitely a reasonable result. And again, think of the progress over like the past 6 months or past year. This would have been not possible a year ago from like the top
models, which is just crazy to think about. I like to think I've played a part in the skate games getting better. All right, so one of the tests that I had run that I think I did this last
night and I never really looked at it is to create the 80s wrestling game using Blender for the assets and then GDAU for the game. So, I asked it for a summary and how do I run it? Because it didn't
really just cleanly give me that at the end. It said a neon 1980s wrestling game. Okay. Blah blah blah. And then here's how we run it. Good. So then I think, can I just open this in terminal
and run GDAU? I'm still kind of learning how to use this software. Very good. [snorts] Okay. All right. Interesting. First and foremost, I'm going to notice a couple of things. Can I Oh, good. All
right. We do have music. It is pretty simple. And I will say some of these models do make this look more of like a sumo wrestling simulator, which is an interesting idea. I'm going to turn
music off for now. Luna Starlight, Bull Brother, Macho Malone. I think we're probably going to have to go with Macho Malone just because I like the name the most. All right, let's see. So,
[laughter] they do it. It's like, you know, they're simple. This definitely gives more vibes of like a Flash model if that is true. And I I believe it probably is though. I'm going to say
like it's not actually bad. It's just simple. Do they have any form of like rigged movement or anything? It's hard to tell. And also, okay, I just wanted to make
sure there was Okay, so there are there is game logic circuit breaker. I It seems like Okay, there's a timeout for that. >> [laughter]
>> This is simple, but I I honestly think as like a foundational point, this isn't bad. And the speed of this model's decent. So, keep in mind that this is Yeah, this is not like a Fable result.
Uh, excuse me, an Opus 55 result, I should say. But for a potential Flash model, this would definitely be acceptable as it did create these assets itself. The models are
doughy and it kind of works. It's simple, but Oh, okay. Free play. 100 HP. Wait a second. I saw one more thing. Let's just try someone else. I E is to interact
with the folding chair. Okay, I think we have to like Where did it go? I saw something. Press E to throw. Oh, that's cool. So, you pick it up and then it disappears cuz
like you're essentially holding it. [sighs] I'm going to get to the chair. Okay. Folding chair acquired. chair smash. Okay, so it it does actually let you. It's just it's
just kind of hard to like see it. All right, overall this was simple, but for a Blender and GDAU result, it's not actually it's not bad. It's just simple and it did definitely have the 80s
aesthetic to it. And that was done through open code as well, not Pi. All right, next up, let's try the Subway FPS result. This is where it needs to use 3JS. It's just going to run in the
browser and the subsequent waves need to be delivered to us via train. All right, so apparently our Subway game has completed and just to see context length. So that was 13.4% of the million
token context window right there. And it seems like it did autoplay it. Oh, okay. So it seems like it did autoplay it. Okay, interesting. We have a lot of the kind of neon aesthetic right here. And
we have last service quarantine district 0317 a.m. Okay, so that's not the correct time in our local, but we'll forgive that. Let me make sure. Okay, sound is on. All right, this is one
where I probably should have asked for the brightness slider, but overall, okay, we have different weapons and we can rightclick to zoom. I don't that's not actually included in the prompt. So,
are there bullet holes in the >> [laughter] >> Okay. Well, there are shell casings. Okay. The train is coming in. I Okay, good. Let me go back to the
pistol. All right. This is honestly again, it's kind of reminds me of like the skate game where really this is this is not bad at all. This still I feel like these
are like GPT6 like probably Luna results actually. I don't know if Luna would do this well. And look at the glowing health pack. I'm actually pretty satisfied with this. Okay, there are
some glitches like the train um doors were not necessarily in the correct like location. The time is also changing. So that's last service 318 platform 6. Let's go back to this just like
ridiculous weapon right here. And they are supposed to get more intense as the waves progress. So [laughter] this weapon is just this is just not
right. Oh, okay. It's like a cheat code. [snorts] And we have health pickups. The lighting effects in the train were not bad. We
can't go uh onto the tracks. Like just seeing the way the headlights come in. That's not bad. Try our best. Yep. Okay, cool. And then some of them are physically larger as well. Like that one
there. And it seems like that one may be a bit tougher. Ah, you were overrun. All right. Overall, you know what? This was actually competent. It was like it had almost these have like neon like
cartoonish effects to them. I'm noticing a pretty um consistent vibe throughout some of these generations. And I personally kind of like the aesthetics of these. So that could also influence
some of my judgment. But especially keeping in mind if this is like a flash form of model, then I'm really like I very much want to see what this is. So next up from within open code just to
try some different harnesses as well. I'm giving this the watch website test where it has to create the beautiful watch cinematic hero section panning over the watch, the exploded view, and
then of course the special edition dubbed the Behan that uses a photo included in the repository or folder of me. So that has to be placed in the special watch on the face and we'll see
what we get. All right, we have our watch website result completed including of course the Bejian special edition just in a single HTML file. So let's see what we get. Okay.
No, you put the special edition in the hero. Normally that doesn't happen. The plus side is it's a good way to butter me up. The downside is it's
showing a few oddities. Now, a couple of things just I'm This was a bad idea because I'm just now distracted by this panning effect right here. It's got the kind of traditional
style that we normally see with this with the black and the white text and the gold text aesthetic. I'll say overall the model is not badly done here. However, it's definitely less
detailed than what we're lately been seeing from definitely some Frontier models. I'll say that. Oh, okay. I didn't notice that the um I genuinely didn't initially notice the hands are a
bit funky because the background of the watch has just totally distracted me. It does say the Bjan right there. All right, let's take a look down. We do have some leather straps and then some
other things. I feel like these are accessories on the table. Let's just scroll down right here. Okay, a quiet object for loud lives. We have a I don't quite know what that is. Nothing
extra. Nothing hidden except that piece of text there in the background. All right. And we have the Bjan. Good. So, we do have the expanded view. Ah, that's what those
were. Okay. So, the things sticking out were the pieces of the inner movement and things like that. So, okay. Overall, this is definitely giving us more flash vibes in terms of the testing. However,
again, if it's a flash model, then that's not as big of an issue. However, these these I'm not super keen on. I mean, this one obviously I am, but I think that this could definitely benefit
from some additional pushing. So, here's what I want to do. Let's get to it, bro. All right. So, in 8 and 1/2 minutes, it said, "You were right. The collection visuals were still
CSS. They're now real 3JS watches." And basically, like, it's just better. So, let's take a look and see for ourselves. Okay, here's what I noticed. It fixed the hands, which is good. We still do
see some of the movement pieces kind of poking up through the face, but the strap also kind of looks a little better. I feel like it's done some things better here, but without seeing
the source directly side by side, it's a little more difficult to tell. But that's actually [snorts] a special edition watch, though. Okay, we still have kind of like the the
jankness right there, the SVG version. But wasn't this invisible or was it the text here? Okay, it was the text here. Let's see how the exploded view looks now. Oh, cool. All right,
assembled and then fully exploded view. I think it just has the same like better features from the hero section one where the hands are actually properly oriented and then also it has some movement gears
in there. Kind of simple. Now before it was just as good. This is what the big problem was. Actually I quite like this color scheme. However, the big issue was that
this was just showing these as SVGs which was absolutely unacceptable. Now with that. Okay. So the button doesn't do anything unless it does and I'm not noticing it. No. Okay. So
sometimes when we have that it puts them in the hero section so we can see them better. However, it's just opting to go with the bejan in the hero section which I can't blame it for. These are
definitely improved. The actual faces on these are kind of nice. they have like some individual patterns and overall it's still simpler but it did fix some of the main issues we were noticing like
these being SVGs and then the hands also not being great. So overall this is an acceptable result and again a lot of this is going to depend on what class of model this is which still we don't 100%
know. So the last thing we're going to do is a test that I came up with randomly that I figured would almost test their ability to like optimize but also has a visual component. I hope this
makes sense. I do have a visual guide to assist us if need be. So, on an old flip phone, perhaps some of you are not aware. Basically, if you wanted to text, you had to press the number button like
a couple of times to get to the specific letter because there weren't enough individual keys here where one equaled each letter. So, each like number right here basically had three keys. So, if I
wanted to press L, I'd have to press five three times. So, J, K, and L or something of this sort. Now, the prompt is, design a better way to text message using an old flip phone where you had to
press each number key a few times to select a letter using the same number only keypad limitations. Try to think of a better, more efficient way to be able to text with those keys and produce the
result as a 3JS simulation. So, I view this less as like a 3D visual result and more of almost a puzzle and we'll see what we get just as a prompt to test their like thinking and novel design
maybe. All right, we have our result right here and it seems kind of interesting because this is also something we've never tested before. So, let's take a second and actually see how
it thought about this. a predictive 12key language, two to nine. One press per letter using T9. Cycle predicted words. Commit prediction and add a space. Cycle common punctuation. Hold
pound. Send the message. No multitap waiting. So, it almost seems to have gone about this instead of like optimizing how the keys do the letters, saying like, okay, these are the most
commonly used letters in the English language, so they're likely going to be placed in like a easier way to access them. It seems like it enabled some form of predictive texting, which is
interesting. So, let's see what we get. Okay, that's actually a that's a pretty cool UI. The model's not bad at all, but it's uh it's just kind of dark, but I guess the focus here is definitely more
on drag to explore. Click the keys to use your keyboard key form. Nice little logo. Start typing below. I have to figure out how to use this.
Effort, not attention. Let's just Oh, it's actually showing that on the screen of the phone, which is cool. Okay, so it's doing predictive text. Tap a letter at once. Tap the star to
cycle or the S3. Prediction three or four. >> [laughter] >> And it has some metrics and things here. Keep in mind this is not as easy to like
understand off the bat as some of our other results where it's pretty easy to like oh pass versus fail. So there's also an approach on this. But also like if you like this style of prompt, let me
know cuz I'm looking for things to also add in that are novel and definitely not in the training data. The rooftop has the best view tonight. Okay, forget about Well, how would we respond to
that? like text 063 028. All right. So 063 0 28 meet at Interesting. Then 03 4448
3 444 8 meet at 8. Oh, meet me at 8. I think maybe we It should have said meet me at 8. Let's try that one more time. 063
me. Oh, okay. Oh, nope. That's not It's okay. I'm starting to kind of understand it. Let's just send the message. And we have to hold it down for the
message to be sent. That's actually kind of cool. Just Oh, just Oh, okay. I'm just going to try to send the message again. I I'm starting to understand this. Okay. All right. Check
this out. So, to start typing, we can type a letter and that will toggle through the predictions is pressing it multiple times. Then, if you press zero, it basically will swap between words and
numbers. So, if we press that again and go back to five, it goes back to the like, well, I think I had it and then I became confused. However, it started to make
sense. And then we can send that and we see it actually does show up there somewhere. I think it's harder to see there than it is. All right. Overall, I don't know what to
say. This was kind of cool and something I wanted to do cuz it's new. So, overall, that's going to conclude a quicker look at Space Bunny Alpha. I don't always test stealth models, but
when they seem kind of interesting, I do. And a model that has multimodal in, so even video, we didn't get to testing that, but that's a more unique thing to have video in. It has a million context
length and it is quite speedy. Now, again, if this is a flash model, that is probably going to be kind of exciting. And we have to keep in mind that without knowing strictly what this is, how large
it is, like what class of model this is designed to compete in, it is really quite difficult to judge the results based off of other models we've seen. With that said, I do have some insight
that I do believe it's a flash model. And I think for a flash model, it performed quite competently. It was fast and I noticed it did a good job like working agentically on the system. I
guess could be said troubleshooting its results and things of this sort. Additionally, it seems to have like a style that it imbued throughout every single result of almost like a colorful
like neon like synth wave kind of style. Even from what we saw right here in the FPS, which was interesting, especially this weapon [laughter] that we had right there was extremely aggro. Everything
just seemed like kind of clean and it did really like this UI style right here. So, that's something I'm going to notice is there was like a common thread of aesthetic woven through this perhaps
to say. Overall, I suppose that's probably going to conclude it. I'm not going to do as much of a traditional results overview because one, we don't know specifically what this is, and two,
there weren't as many tests run with this. When we do test these stealth models, I kind of just like to do it a bit simpler and maybe throw in some new things like the phone test right here.
So that is going to conclude our first look and test of Space Bunny Alpha which can be used through open code right now. It's also available on open router. So that's kind of how we were using it
today. I am very very interested to see what this ends up being. So with that, if you have any questions, please feel free to leave them in the comments and thanks for watching.