Is that a pigeon? I think it might have been.

Today we're going to be testing Meta's new

updated Muse Spark, which is Muse Spark 1.1.

Now I tested the original one when it came out,

and I do have to say I was actually rather impressed with it.

It wasn't the most potent

model on the block, but it was rather fun to play with,

and it did end up making some enjoyable

games. Right now I am using this from within open code,

and I have given this a prompt that

I ran with Fable 5 to create a game called Castle Time,

but there's a lot of additional niche detail

that this needs to include, and we'll get into that a bit later,

but I do just want to start out by

taking a quick look at the introduction post.

Of course, this does seem like a significant leap in

capability over its predecessor based off of what we see in these benchmark charts right here. Now,

for the purpose of this, I am going to, for this moment,

ignore the other models here, but look at

the base one, and look at the jumping capability.

On Deep SWE 1.1, it went from a 10 to a 53.3.

Sweebench Pro went up, not as much of a significant leap,

but still, this does seem like a proper jump

over its predecessor, and for that I am excited.

Now, because this is compared to other models like

3.1 Pro High, which is old now, Opus 4.8, and GPT-55,

I'm going to test this properly as I would

were I testing any one of these.

This will be getting no breaks or anything of the sort. This

is a new entry into the Frontier state-of-the-art model space.

It will be very interesting as well

to see how it compares to Grok 4.5.

That is not included in the benchmark charts here, but

doesn't matter. We can tell on vibes as well and

things of the sort. So, before we go further,

please do feel free to subscribe as I do want that 100k plaque,

and let's just take a quick look at some

of the additional information in this announcement blog post,

although to be honest, it's a fairly light

blog post in terms of the model announcement.

It doesn't tell us a huge amount about it. It just

basically talks a little about it, shows some demonstration,

and some benchmarks. So, the feel

for this will definitely be coming mostly from the testing.

They do have this meta internal coding

bench where this is showing like just under Opus 4.

8. Now, let's take a look at some of the interesting

things for at least development or API use of this model.

It is price-wise very competitive,

though I do need to say that with the caveat that I don't know where this sits in terms of what models

it's competitive with from OpenAI or from Anthropic.

So, these prices may change depending

on whether or not it's comparable to like Sonnet or Opus or something.

So, it's $1.25 per million in

and $4.25 per million output tokens.

They have some additional info here just pertaining to like

using this as a developer. However,

something interesting is the context window for this is

listed at a bit over a million tokens.

I don't have any specific tests that will really push its

context window to that degree and see how well it does,

but just having that over a million is pretty

exciting, I would say. All right. So,

from within the web playground, I am going to be doing the

browser OS just so

we can run somethi

ng a little more intricate at the same time in OpenCode, which is

that Fable 5 game test that I have a result from Fable 5.

It was a lot of fun to play. All right. So,

in just under 15 minutes, Castle Time has been created.

It is running apparently as well,

so we can just pop right to there. Port 8087.

I'm excited. All right. So, let's check out the Muse

Spark 1.1 Castle Time. Okay. I want to make sure that we do have sound.

The speaker is on. It should

be giving us some sound. All right. Wake up, Chrono.

I found it. The scroll your thesis was about. The one

they say knows how to fold light. Okay. So,

we have to continue. Let's look around in the bedroom.

Okay. Not bad. It has those vertical weird, like,

LEDs that you sometimes see. All right. I'm just

going to... A, E to continue.

It's designed to work with a

little Bluetooth controller, but, um, keyboard

fallback as well. So, like,

it doesn't seem like the ke

yboard fallback is working. That's nice. Why is it...

Investigate the buzzing device on your desk. So,

we need to go over to it.

keyboard. E is supposed to work. Wake up, Chrono.

I found it. The scroll your thesis was about.

Where did it get this voice? Where did they get this voice?

All right. I'm hitting E. You...

Okay. I have a few choice words. Actually,

I'm just going to... I'm going to be calm.

I'm just giving it some follow-up, like,

the keyboard fallback is not working control-wise.

I also, like, what is that voice?

Let me take a look at some of these assets. Okay. Wow.

It's genuinely a single HTML file. The other... I.

.. The prompt doesn't say it needs to be.

The other models have done this in significantly more files, so we'll see.

All right. Supposedly,

it fixed the keyboard not working properly, so...

Wake up, Chrono. I found it. The scroll your thesis was about.

The one they say knows how to fold light.

Okay. Just please let me, like...

E.

Okay.

Why can't...

I interact with anything?

A.

E.

Click.

Space.

I even have the controller hooked in, and...

And it's not doing anything.

Okay.

Okay.

Like...

Like...

Why do the keyboard...

Okay.

You know what?

We're going to come back to this.

In the meantime, let's do our browser OS test.

It's not funny. I'm just...

I don't...

So here's our browser OS test.

Okay.

It did it in 549 lines of code, and

it started out by spoiling the GTA game, which is infuriating.

Though I'm going to say the rest of this is quite acceptable.

Clock is correct.

I won't spend as much time on these.

I know they're a little outdated.

It didn't include a right click, which is...

Surprising.

Breach.

Off or on.

That's probably a special feature, so we'll save that for last.

Alright.

Let's check our start bar.

Okay.

Sunset Dune.

Deep Void.

Living City Live.

That's a sick live desktop background.

It's like a top-down view of the city.

Or Matrix Code, which is...

Just...

I don't know.

Okay.

It's a fourth wall break engine.

The special feature.

Very creative.

Let's just look through the apps.

I'm going to change the desktop background because

it is a little distracting here.

We'll just...

I like Deep Void.

I think that's cool.

Files.

BreachProtocol.exe.

TimeShards.

Drag a file to GTA window to spawn object.

Okay.

And these don't open, but...

Interesting interactivity.

I think we'll go back to that.

It seems very, very focused on the special feature here.

I like that.

It's like one off on the bottom, but I like it.

Nonetheless, let's try Pixel Paint.

Can we resize it and still have the canvas change?

Good.

Not bad.

No way to save it, but...

It was acceptable.

Block City GTA.

We'll save that for last now.

Nova Drift.

Well...

I can't actually...

Like...

I'm spam clicking the key to move, and it...

It's impossible to not hit these.

Space is to break.

Okay.

W is boost.

Okay.

We did speed up the...

This is a cool game.

It's just...

It's...

It was impossible to avoid the object.

Okay.

E is to enter or exit a car.

Drop a file to spawn ramp.

This is where...

It seemed like the file system being open here.

So we can...

That's actually kind of sick.

Good.

Good.

Um...

Arrow keys?

Oh.

Shift for turbo.

The problem is we can't actually go forward when in the car.

Now the keyboard and mouse have stopped...

Um...

Affecting the movement.

This thing seems to have a difficult time just mapping keys to game events, which is...

Curious.

Okay.

Can we...

All right.

Good.

Shift is for turbo.

Good.

There are mesh colliders.

All right.

Um...

Games can now leak to desktop.

Okay.

Maybe...

That's cool.

Look.

When we actually fail in the game, it affects the desktop.

That's...

That is neat.

I like that.

It's very creative.

And the way it actually changes the colors is cool as well.

All right.

That's a very interesting result.

All right.

Now we have to go back and see what exactly is stopping the keyboard.

I'm going to clear cache.

Let's try again.

Wake up, Crono.

I found it.

The scroll your thesis was about.

Okay.

The one they say knows how to fold light.

All right.

So it's not working and...

All right.

So I'm telling it to just fix the code.

Is it going to say all set, master?

I'd accidentally closed the open code session.

It may have been a Freudian close, but I just...

We'll just see because

it was trying to like run it again and check it.

And I had...

Okay.

Okay.

It said the line.

Good.

So now hypothetically, if we go back in here and

then just run the same thing again,

and then go to the URL, it will hopefully have proper keyboard.

And I have to listen to this voice.

Wake up, Crono.

I found it.

The scroll your thesis was about.

The one they say knows how to fold light.

We'll just pretend like this is our first time.

Like, oh, okay.

Very cool.

So now...

E.

There's...

Something's like...

Spent.

Spent.

I've literally spent like $6 to just get like...

Like kicked...

I need to ask Claude if I'm...

If I'm...

All right.

So we're having Claude fix this game because...

I gave the meta model a few chances.

And Claude gave us actually a valuable assessment of the issue here.

So any click permanently jams the laser on.

Okay.

That's fine.

Dialog advanced is wired four times over plus an 80 millisecond polling loop.

So the intent is clearly there and it does fire.

Continue skips lines unpredictably.

Here's, I think, what we're facing.

E to interact only does something when

you're basically on top of the object.

So you need to be within 1.8 units.

And I'm unable to seemingly get within there.

So I'm just asking Claude to just fix this and

change it so we don't need to be that close.

And then we'll take a look at it.

Just keeping that in mind.

The game is still Muse Spark.

Wake up, Crono.

I found it.

The scroll your thesis was about.

The one they say knows how to fold light.

So we should be able to...

Thank you.

Interact.

Okay.

So our character isn't speaking, but that's okay.

It could be a mismatch with the voice or something.

It was.

On October 10th, 1283.

Carnarvon Castle, North Wales.

53.1393N.

Minus 4.2768WA Da Vinci type polymath was imprisoned in the dungeon.

He had it.

But tonight, a trebuchet misfire will set the south wall ablaze.

Everything burns by midnight.

If you want it, you leave now.

You didn't wash your laundry though.

Good luck.

A bit of humor.

Okay.

Your Motorola time port on the desk is buzzing.

It has a temporal antenna now.

Interact with it.

Nice.

Good, good, good.

Okay.

Finally.

Oh, initiate jump.

It's interactive.

Don't think about socks.

You forgot socks.

Demon.

His clothes.

Shine like devil skin.

Get him.

It didn't.

It didn't.

Wake up, Crono.

Okay.

But tonight, a trebuchet misfire will set the south wall ablaze.

Everything burns by midnight.

Are you kidding me?

You forgot socks.

It didn't.

Demon.

His clothes.

Shine like devil skin.

All right.

You know what?

I'm sorry.

This test is a little too much for this model.

And that's okay.

It's a test that I've only run so far with Fable.

We'll just keep going.

All right.

So I'm giving it the self-contained C++ skateboard sim test.

The California boardwalk aesthetic version.

And we'll see how it does.

All right.

Cool.

So it's just checking what specific things are available on this system for it to build

this with.

And it does not seem like it's going to try to use Raylib.

So we won't have to redirect it.

Some folks say, why not Raylib?

It makes it simpler for the model to perform.

It is allowed for significantly smaller models like Gemma 412b or something.

But for larger state-of-the-art models that cost a bit of money to use,

I like to see their

more performant capability without using Raylib.

All right.

The code's been written.

It's right at 1600 lines.

So it's going to verify it builds and then package the final thing.

So we may see our first glimpse of it depending on what it does here for QA.

Okay.

Now we got some errors when it was trying to build it.

Only a few.

So it should likely fix these, which it's doing now.

And then it will rerun to see if it builds.

All right.

And we have our result.

So let's take a look at it.

And it is done.

I know folks are saying, let it finish.

It was just doing like, that's not half bad.

That's really not half bad.

Especially considering the issues we noticed with the game that like took years off my life

to try to get working.

Not bad.

The shirt changes colors when we're grinding on the rail.

I noticed the shirt changing colors is a common thing in this test.

I have no idea why, but I noticed it a lot with open weight models.

All right.

Kickflip animations.

Not bad.

All right.

Well, I think we hit the NPC and it actually made us bail,

which was, that was cool.

Good.

I like that.

All right.

What else do we have?

A lot of tricks here.

B is a 360.

Oh, just like that.

All right.

All right.

Let's take a look at the environment.

Let's go back.

We do have, the palm trees are actually pretty nicely drawn.

I'm going to say the water effect is acceptable.

The storefronts are acceptable as well.

Nice.

Dude, watch it.

And it even put like clever little things there.

I'm actually, I'm okay with this.

It's not like state of the art frontier.

This is definitely closer to like something you'd get with GLM 5.

2, Kimi K 2.7 than any of the high end GPTs.

But nonetheless, it's, it's solid.

So from the web chat here, I'm giving it the beautiful front end test.

And this is the one where it needs to create a 3D model of a luxury watch.

Put it in the hero section of this website with a nice cinematic camera panning shot.

And then the pricing cards for two models of watches should also

have 3D models that it creates.

So this should just make a high end front end with a nice 3D modeling and

camera panning thing.

So here's our front end design test, the watch website in specific.

Okay.

Not quite.

I actually don't mind the font it's used and the watch co section.

It's just the models, unfortunately,

not up to par with what we've seen from other models.

Mainly the missing hands and things like that.

Maybe it's an LED watch if we were to give it the benefit of the doubt.

Okay.

So I'm a bit troubled, but

these watches are signific

antly cheaper than ones we've seen on other sites.

So maybe that's a reflection of the rendering here.

Let's give it some reinforcement.

No, I, I, I don't want that.

Stop it.

Okay.

So I responded alert expecting it to be like, I'm sorry.

Like, what do you mean?

It just went ahead and

thought that meant add an alert bar for the summer 26 drop.

So I had to stop it and say, sorry, I meant garbage alert.

And we'll see now if it tries to improve the result.

Got it.

You hit the garbage alert.

Let's take it out.

Nice.

The last one was too low poly and too much wood.

All right.

T-shot grade now.

I like it.

It understood intent.

So here's the second try of our watch website.

Okay.

It's still loading.

It just, it didn't make much of a difference.

It's slightly actually different.

It's just unfortunately not really.

Next up, I'm going to try a multimodal coding test just to see how it does.

I see the ZX Spectrum photos.

How dare you?

That is my own specific design.

Thank you very much.

So it's tasked with creating a 3D replica of this laptop that can be opened and explored.

We also need to have a functional keyboard where we can press things in the keyboard that it 3D models and

they will appear on screen.

So this is more of a test of its capability of seeing something and

then transposing it into 3D based off of a decent amount of reference photos.

All right.

Let's take a look at our photo to 3D model slash interactive website test.

Okay.

You know what?

That's, that's not bad.

It's not the best one I've seen, but

oftentimes models will invert the keyboard for some reason.

This hasn't done that.

Let's check the open close capability.

Okay.

This actually, it's not perfect, but

it's definitely better than some of

the open weight models I've seen do in this task.

Toggle RGB.

Toggle RGB.

Cool.

Reset cam.

Now, we should be able to type on the screen.

The problem is,

the screen is either hidden behind something or it's just not working.

But the key press events don't look bad.

I will say, those are decently solid effects there.

The close animation is not bad either.

Interestingly, no model really has gotten the hinge to be in the correct location,

which would be over here.

There's definitely some issues here, but lid open 160.

Oh, that's just frames per second, like, from this scene.

It's passable.

So, next up, I'm going to give this a 3D model test where it needs to create a 3D model of a V8 engine for my remote control car.

I have a 280 sized motor that I'd like to put inside of this engine model.

So, I wanted to make it something that looks realistic and

also is something that can be 3D printed.

We'll see how this does.

Interesting.

It just wanted to find out about the system.

And it's looking up motor specifications.

Good.

Using standard 280 specs.

So, my camera wasn't recording for the 3D model test,

which, you know, it does happen sometimes.

Basically, overall, we'll look at this once more.

It actually did a solid job of making a replication of a V8 engine.

Some of the things it included were actually very interesting and

levels of detail that I've not seen in a lot of other models when

doing this prompt.

Unfortunately, the big issue,

as we're going to see here if we open the SCAD file,

is that the hole for this engine is not deep enough to fit a 280 motor.

And even if it wasn't intended to go all the way in there,

the problem is this base here just cuts into the cylinder.

So, it would be physically impossible to install that in here.

I did also tell it to optimize this so

it could actually be 3D printed.

And unfortunately,

there's just a bit too much odd geometry everywhere.

And if we open this in a 3D slicer program,

we'll be able to see a bit better in reference to that.

So, this would be very difficult to print.

Just with, like, gaps like that, you would need...

I mean, it wouldn't be horrible to print, but

it would need a lot of support.

And then we have an issue like the air cleaner here on top of what would be the carb is just floating.

The valve covers are also kind of floating.

The connection's not great to the Vs.

But the things that it did do that were good that are just weird is the valve covers actually have,

like, individual hex bolts that show, like, them being bolted on.

And they're really quite well done.

Now that I look more at that, though, yeah.

I mean, that's cool.

And then the other thing it did,

it put an oil cap on one side of

the valve cover, which is kind of how it would be in real life.

So, just weird detail that it included like that that's not often seen in other models.

And then, unfortunately, it just wasn't altogether competently.

But it does show some level of promise.

And on first glance, just from, like, zoomed out,

it's like, whoa, this is really quite good.

So, that was my initial reaction that I failed to capture on camera because

I was not recording.

Nonetheless, we're going to move on now to a different task.

And that is going to be one that I unfortunately did not get.

I tried doing this in my GPT-5-6 test, and

it was the last thing we ran, and it just

froze up, and I was never able to play it.

So, I've just taken a screenshot of that prompt.

I had meta OCR it from this web chat interface, but

I will have it be built through open code.

I'm going to start it in plan mode here, because why not?

And we'll see how it differs, if it differs.

Got it.

Street eat.

You are BigFist McGee in a pastel low-poly city block.

Okay.

That name was not included in the prompt.

So, it came up with that on its own.

I did see some folks talk about how this could be,

like, funny to talk to.

We'll try that later.

Okay.

That was a good plan.

Open questions.

Now, the original Muse Spark, when I tested it,

it made a very fun Steve the PC Repairman offic

e game where you just go around and smash PCs and beat up NPCs.

So, it'll be interesting to see how this does.

Alright.

Let's check out Street Yeet.

So far, I like what I see.

One fist to yeet them all.

Hold mouse for a mega charge.

Look for golden objects.

That's a bonus Yeet.

Try to get the buses.

I dare you.

Again, we can punch pigeons, which was included in the GPT-5-6 game.

It's not included in the original prompt.

This is not bad.

Well, let's, uh...

Oh, yeah.

Alright.

The pedestrians are not quite right, but...

Hey.

Why can't I...

Okay.

Okay.

The cartoon style and stuff.

Alright.

It said try to yeet the bus, so let's just...

Let's just try this.

It didn't go as far...

Okay.

Now, that's actually...

And it has sound effects when we hit it.

Cool.

We're supposed to collect these, but I don't see how.

Let's just yeet this car.

Make it delete.

Alright.

I can't yeet the pedestrians.

Oh.

Oh.

Oh, I wish they looked more humanoid.

Is that a...

Pigeon?

I think it might have been.

Well, birds are meant to fly.

I can yeet everything here.

Elam, yeetus?

Why is it...

It's models.

It's just...

Okay.

The model's fun.

The Muse Spark.

Even the first one.

It's good at making fun or funny games.

Alright.

Let's try some multimodal coding.

Also, I'm gonna turn off this just rhythmic pounding noise.

It's questionable.

Yeah, I see it.

You're 10 yeets deep.

13,000 chaos.

14, 19 meters.

Far this nice.

I also see some issues.

Okay.

I wanted to include that the humanoid, like the NPCs,

are also garbage.

But nonetheless, we're gonna just let it ride.

Alright.

Let's check out our hypothetically improved Street E game.

I'm not really thinking.

Let's just try some creative writing.

So I'm giving it this photo,

which you may be familiar with if you watch the channel.

So thank you.

So I'm saying this is the cover for an upcoming psycho thriller romance novel.

Give me the title and back cover summary of the book.

And we'll see what it says.

Alright.

Alright.

Alright.

Title.

The way he holds me.

Let's zoom in a little.

Tagline on the front.

He'd do anything to keep her smiling.

Here's the back cover summary.

Everyone said Anna was still lucky.

When Mark asked her to leave the city and

move into his beautiful, crumbling Victorian home

three hours from anyone she knows, it was romantic.

When he deleted her social media because

you don't need anyone else, it was devoted.

When he holds her for photos, his smile is perfect.

Hers never is.

In the picture on their mantle, he is beaming.

His arm around her shoulder.

To their friends, to their neighbors, to the camera, it's love.

Anna knows better.

It's a warning.

Stay still.

Smile.

Don't pull away.

She has been planning to leave for months.

She has a bag packed in the attic and

a bus ticket hidden in a loose floorboard.

But Mark has been planning too.

He knows she's unhappy.

He knows she's been lying.

And he has no intention of letting his wife ruin their perfect love story.

A claustrophobic, slow burn,

psychological thriller about performative love control and

what happens when

one person who says they'll never leave you means it.

Alright.

That was...

That was good.

So I'm asking as a follow-up, give me a TLDR of the whole story.

Anna marries Mark, the perfect you're so lucky guy,

and moves into his isolated home.

Very quickly, the romance turns coercive control.

He isolates her.

Okay, we already kind of did this.

She secretly plans to leave.

While packing, she finds out Mark's first girlfriend didn't just leave him and move away

like he said.

She disappeared and the police never found her.

On the night, Anna finally tries to run.

Mark catches her with her go-bag and

forces her to sit for one last photo.

The photo from the cover.

He tells her no one will believe her if she does leave.

She pretends to give in.

That night, she uses all the things he's taught her about how he watches and

controls everything

to turn it against him.

Records his confession.

Locks him in the old root cellar and calls the police.

It ends with Anna back in the city a year later, living alone.

She gets her mail and

there's a new framed photo inside of her sleeping in her new apartment

taken from her outside window.

Alright.

Alright.

This definitely was something.

So next up, we're going to be doing the 3D city block that is shown in different time

periods.

We'll see how it does.

This is one that I've more recently introduced as a test and

it's always kind of fun to see

what the models decide to include, especially for the future of 2055.

So we'll see how it does.

Alright.

In 4 minutes and 8 seconds,

we've received the result for the city timeline test.

Very, very quick.

Keep in mind, none of these are instructing it that it needs to do these in single scripts.

Okay.

It started us in 1985.

That's definitely noticeable.

Sunset VHS.

Alright.

We have a button there.

I'm not 100% sure what these do as of this second, but that's okay.

Oh, it froze.

Alright.

We'll just refresh.

Okay, cool.

The cars are moving.

There's a car in the store.

It's possible there was an accident.

Let's go to 1945.

Kind of cool.

Look at it.

It got more grainy.

And the folks like the...

It's almost grainy to a fault, but

the people's clothing is changed to reflect.

Even the car that's driven into the building looks a little older.

The streetlights have changed.

Let's do 1965.

Much happier.

Very, very mid-century modern looking colors on the buildings.

And the car does have fins.

Good, good, good.

Okay, this one's going backwards.

That's alright.

The people's outfits are more colorful and happy.

Drink Atomic.

Mod Boutique.

And the streetlights have changed to mid-century modern shapes as well.

Kind of cool.

Alright.

This is still...

It's basic.

This doesn't have the level of depth and

intricacy that the other state-of-the-art models do.

Not at all.

But it's still...

Okay, bland, like corporate 2005.

Everything's a little more toned down.

2025.

Okay, we have a sound.

It's still like...

Solar panels on the roof.

It's got some understanding.

It's just not at the level of the other models currently that it's being benched against.

Alright, we have hover vehicles.

Everything's kind of neon and glowing, which I personally like.

It's just interesting how a lot of models converge into this.

Almost kind of like 85 inspired, but in 2055.

So, some sounds...

Chrono Block Manhattan.

After Light.

Present Tense.

Analog Sunset.

Neon Decade.

Dream of Tomorrow.

And...

Ashes and Victory.

Interesting.

Alright, the last test I'm going to do is from the web chat interface right here.

It is using this model.

This is the suggested temperature from the docks.

On X-High, this is going to be the unified,

beautiful subway scene, as well as turn it into an FPS.

Just to see how it does on one final, common game test.

Alright, let's check out our subway FPS result.

I'm giving it a follow-up here that it will hopefully understand to fix this.

Alright, so here's our hypothetically fixed subway...

Subway...

The same problems occur.

It just changed the theme to, like, light mode from...

From dark mode and...

Oh, wow.

Okay.

Alright.

Third time's the charm.

Good.

Very good.

This is a big station.

Nice.

Now the big test.

No, but we get some effects, like, of the ammunition hitting,

so I'll take that.

This is one of the few ones that actually shows,

like, the subway tunnel instead of just being bl

ocked off, which is a curious...

Flat line.

Alright.

Well, that's going to conclude our first look and

test of Muse Spark 1.1.

I have to be totally honest that I don't think this model is ready for prime time.

They gave me $20.1.

They gave me $20.00 in free credits.

They give everyone that if you put a billing thing in.

The entirety of this testing used $11.84.

It's, um...

Again, I don't think it's ready for prime time.

It made a lot of just...

The results just weren't that good.

But some things showed promise.

And things that it did that included that, like,

I don't see commonly in these tests.

Like, the subway actually had a view beyond where the subway car would go.

Most models just block it off.

The 3D model test actually having individual bolts for the valve covers.

The oil cap in the correct side.

But I have to say, for the pricing of this model,

I don't really see the value in this.

I think Composer 2.5 is cheaper and better.

And, yeah.

It's an attempt at improvement from a player who hasn't been involved in the game for a while.

And that's good to see.

I would be interested to see how this progresses.

Maybe over the summer.

Maybe into fall.

But as of right now,

I have to be honest that I am disappointed with this model.

Especially considering the price.

Make of that what you will.

Overall, I wanted to test this because I was very excited to test it.

And it's, again, it's an early entry.

Or improvement.

So, with that, if you have any questions,

please feel free to leave them in the comments.

And thanks for watching.

Is that a pigeon? I think it might have been.

Today we're going to be testing Meta's new

updated Muse Spark, which is Muse Spark 1.1.

Now I tested the original one when it came out,

and I do have to say I was actually rather impressed with it.

It wasn't the most potent

model on the block, but it was rather fun to play with,

and it did end up making some enjoyable

games. Right now I am using this from within open code,

and I have given this a prompt that

I ran with Fable 5 to create a game called Castle Time,

but there's a lot of additional niche detail

that this needs to include, and we'll get into that a bit later,

but I do just want to start out by

taking a quick look at the introduction post.

Of course, this does seem like a significant leap in

capability over its predecessor based off of what we see in these benchmark charts right here. Now,

for the purpose of this, I am going to, for this moment,

ignore the other models here, but look at

the base one, and look at the jumping capability.

On Deep SWE 1.1, it went from a 10 to a 53.3.

Sweebench Pro went up, not as much of a significant leap,

but still, this does seem like a proper jump

over its predecessor, and for that I am excited.

Now, because this is compared to other models like

3.1 Pro High, which is old now, Opus 4.8, and GPT-55,

I'm going to test this properly as I would

were I testing any one of these.

This will be getting no breaks or anything of the sort. This

is a new entry into the Frontier state-of-the-art model space.

It will be very interesting as well

to see how it compares to Grok 4.5.

That is not included in the benchmark charts here, but

doesn't matter. We can tell on vibes as well and

things of the sort. So, before we go further,

please do feel free to subscribe as I do want that 100k plaque,

and let's just take a quick look at some

of the additional information in this announcement blog post,

although to be honest, it's a fairly light

blog post in terms of the model announcement.

It doesn't tell us a huge amount about it. It just

basically talks a little about it, shows some demonstration,

and some benchmarks. So, the feel

for this will definitely be coming mostly from the testing.

They do have this meta internal coding

bench where this is showing like just under Opus 4.

8. Now, let's take a look at some of the interesting

things for at least development or API use of this model.

It is price-wise very competitive,

though I do need to say that with the caveat that I don't know where this sits in terms of what models

it's competitive with from OpenAI or from Anthropic.

So, these prices may change depending

on whether or not it's comparable to like Sonnet or Opus or something.

So, it's $1.25 per million in

and $4.25 per million output tokens.

They have some additional info here just pertaining to like

using this as a developer. However,

something interesting is the context window for this is

listed at a bit over a million tokens.

I don't have any specific tests that will really push its

context window to that degree and see how well it does,

but just having that over a million is pretty

exciting, I would say. All right. So,

from within the web playground, I am going to be doing the

browser OS just so

we can run somethi

ng a little more intricate at the same time in OpenCode, which is

that Fable 5 game test that I have a result from Fable 5.

It was a lot of fun to play. All right. So,

in just under 15 minutes, Castle Time has been created.

It is running apparently as well,

so we can just pop right to there. Port 8087.

I'm excited. All right. So, let's check out the Muse

Spark 1.1 Castle Time. Okay. I want to make sure that we do have sound.

The speaker is on. It should

be giving us some sound. All right. Wake up, Chrono.

I found it. The scroll your thesis was about. The one

they say knows how to fold light. Okay. So,

we have to continue. Let's look around in the bedroom.

Okay. Not bad. It has those vertical weird, like,

LEDs that you sometimes see. All right. I'm just

going to... A, E to continue.

It's designed to work with a

little Bluetooth controller, but, um, keyboard

fallback as well. So, like,

it doesn't seem like the ke

yboard fallback is working. That's nice. Why is it...

Investigate the buzzing device on your desk. So,

we need to go over to it.

keyboard. E is supposed to work. Wake up, Chrono.

I found it. The scroll your thesis was about.

Where did it get this voice? Where did they get this voice?

All right. I'm hitting E. You...

Okay. I have a few choice words. Actually,

I'm just going to... I'm going to be calm.

I'm just giving it some follow-up, like,

the keyboard fallback is not working control-wise.

I also, like, what is that voice?

Let me take a look at some of these assets. Okay. Wow.

It's genuinely a single HTML file. The other... I.

.. The prompt doesn't say it needs to be.

The other models have done this in significantly more files, so we'll see.

All right. Supposedly,

it fixed the keyboard not working properly, so...

Wake up, Chrono. I found it. The scroll your thesis was about.

The one they say knows how to fold light.

Okay. Just please let me, like...

E.

Okay.

Why can't...

I interact with anything?

A.

E.

Click.

Space.

I even have the controller hooked in, and...

And it's not doing anything.

Okay.

Okay.

Like...

Like...

Why do the keyboard...

Okay.

You know what?

We're going to come back to this.

In the meantime, let's do our browser OS test.

It's not funny. I'm just...

I don't...

So here's our browser OS test.

Okay.

It did it in 549 lines of code, and

it started out by spoiling the GTA game, which is infuriating.

Though I'm going to say the rest of this is quite acceptable.

Clock is correct.

I won't spend as much time on these.

I know they're a little outdated.

It didn't include a right click, which is...

Surprising.

Breach.

Off or on.

That's probably a special feature, so we'll save that for last.

Alright.

Let's check our start bar.

Okay.

Sunset Dune.

Deep Void.

Living City Live.

That's a sick live desktop background.

It's like a top-down view of the city.

Or Matrix Code, which is...

Just...

I don't know.

Okay.

It's a fourth wall break engine.

The special feature.

Very creative.

Let's just look through the apps.

I'm going to change the desktop background because

it is a little distracting here.

We'll just...

I like Deep Void.

I think that's cool.

Files.

BreachProtocol.exe.

TimeShards.

Drag a file to GTA window to spawn object.

Okay.

And these don't open, but...

Interesting interactivity.

I think we'll go back to that.

It seems very, very focused on the special feature here.

I like that.

It's like one off on the bottom, but I like it.

Nonetheless, let's try Pixel Paint.

Can we resize it and still have the canvas change?

Good.

Not bad.

No way to save it, but...

It was acceptable.

Block City GTA.

We'll save that for last now.

Nova Drift.

Well...

I can't actually...

Like...

I'm spam clicking the key to move, and it...

It's impossible to not hit these.

Space is to break.

Okay.

W is boost.

Okay.

We did speed up the...

This is a cool game.

It's just...

It's...

It was impossible to avoid the object.

Okay.

E is to enter or exit a car.

Drop a file to spawn ramp.

This is where...

It seemed like the file system being open here.

So we can...

That's actually kind of sick.

Good.

Good.

Um...

Arrow keys?

Oh.

Shift for turbo.

The problem is we can't actually go forward when in the car.

Now the keyboard and mouse have stopped...

Um...

Affecting the movement.

This thing seems to have a difficult time just mapping keys to game events, which is...

Curious.

Okay.

Can we...

All right.

Good.

Shift is for turbo.

Good.

There are mesh colliders.

All right.

Um...

Games can now leak to desktop.

Okay.

Maybe...

That's cool.

Look.

When we actually fail in the game, it affects the desktop.

That's...

That is neat.

I like that.

It's very creative.

And the way it actually changes the colors is cool as well.

All right.

That's a very interesting result.

All right.

Now we have to go back and see what exactly is stopping the keyboard.

I'm going to clear cache.

Let's try again.

Wake up, Crono.

I found it.

The scroll your thesis was about.

Okay.

The one they say knows how to fold light.

All right.

So it's not working and...

All right.

So I'm telling it to just fix the code.

Is it going to say all set, master?

I'd accidentally closed the open code session.

It may have been a Freudian close, but I just...

We'll just see because

it was trying to like run it again and check it.

And I had...

Okay.

Okay.

It said the line.

Good.

So now hypothetically, if we go back in here and

then just run the same thing again,

and then go to the URL, it will hopefully have proper keyboard.

And I have to listen to this voice.

Wake up, Crono.

I found it.

The scroll your thesis was about.

The one they say knows how to fold light.

We'll just pretend like this is our first time.

Like, oh, okay.

Very cool.

So now...

E.

There's...

Something's like...

Spent.

Spent.

I've literally spent like $6 to just get like...

Like kicked...

I need to ask Claude if I'm...

If I'm...

All right.

So we're having Claude fix this game because...

I gave the meta model a few chances.

And Claude gave us actually a valuable assessment of the issue here.

So any click permanently jams the laser on.

Okay.

That's fine.

Dialog advanced is wired four times over plus an 80 millisecond polling loop.

So the intent is clearly there and it does fire.

Continue skips lines unpredictably.

Here's, I think, what we're facing.

E to interact only does something when

you're basically on top of the object.

So you need to be within 1.8 units.

And I'm unable to seemingly get within there.

So I'm just asking Claude to just fix this and

change it so we don't need to be that close.

And then we'll take a look at it.

Just keeping that in mind.

The game is still Muse Spark.

Wake up, Crono.

I found it.

The scroll your thesis was about.

The one they say knows how to fold light.

So we should be able to...

Thank you.

Interact.

Okay.

So our character isn't speaking, but that's okay.

It could be a mismatch with the voice or something.

It was.

On October 10th, 1283.

Carnarvon Castle, North Wales.

53.1393N.

Minus 4.2768WA Da Vinci type polymath was imprisoned in the dungeon.

He had it.

But tonight, a trebuchet misfire will set the south wall ablaze.

Everything burns by midnight.

If you want it, you leave now.

You didn't wash your laundry though.

Good luck.

A bit of humor.

Okay.

Your Motorola time port on the desk is buzzing.

It has a temporal antenna now.

Interact with it.

Nice.

Good, good, good.

Okay.

Finally.

Oh, initiate jump.

It's interactive.

Don't think about socks.

You forgot socks.

Demon.

His clothes.

Shine like devil skin.

Get him.

It didn't.

It didn't.

Wake up, Crono.

Okay.

But tonight, a trebuchet misfire will set the south wall ablaze.

Everything burns by midnight.

Are you kidding me?

You forgot socks.

It didn't.

Demon.

His clothes.

Shine like devil skin.

All right.

You know what?

I'm sorry.

This test is a little too much for this model.

And that's okay.

It's a test that I've only run so far with Fable.

We'll just keep going.

All right.

So I'm giving it the self-contained C++ skateboard sim test.

The California boardwalk aesthetic version.

And we'll see how it does.

All right.

Cool.

So it's just checking what specific things are available on this system for it to build

this with.

And it does not seem like it's going to try to use Raylib.

So we won't have to redirect it.

Some folks say, why not Raylib?

It makes it simpler for the model to perform.

It is allowed for significantly smaller models like Gemma 412b or something.

But for larger state-of-the-art models that cost a bit of money to use,

I like to see their

more performant capability without using Raylib.

All right.

The code's been written.

It's right at 1600 lines.

So it's going to verify it builds and then package the final thing.

So we may see our first glimpse of it depending on what it does here for QA.

Okay.

Now we got some errors when it was trying to build it.

Only a few.

So it should likely fix these, which it's doing now.

And then it will rerun to see if it builds.

All right.

And we have our result.

So let's take a look at it.

And it is done.

I know folks are saying, let it finish.

It was just doing like, that's not half bad.

That's really not half bad.

Especially considering the issues we noticed with the game that like took years off my life

to try to get working.

Not bad.

The shirt changes colors when we're grinding on the rail.

I noticed the shirt changing colors is a common thing in this test.

I have no idea why, but I noticed it a lot with open weight models.

All right.

Kickflip animations.

Not bad.

All right.

Well, I think we hit the NPC and it actually made us bail,

which was, that was cool.

Good.

I like that.

All right.

What else do we have?

A lot of tricks here.

B is a 360.

Oh, just like that.

All right.

All right.

Let's take a look at the environment.

Let's go back.

We do have, the palm trees are actually pretty nicely drawn.

I'm going to say the water effect is acceptable.

The storefronts are acceptable as well.

Nice.

Dude, watch it.

And it even put like clever little things there.

I'm actually, I'm okay with this.

It's not like state of the art frontier.

This is definitely closer to like something you'd get with GLM 5.

2, Kimi K 2.7 than any of the high end GPTs.

But nonetheless, it's, it's solid.

So from the web chat here, I'm giving it the beautiful front end test.

And this is the one where it needs to create a 3D model of a luxury watch.

Put it in the hero section of this website with a nice cinematic camera panning shot.

And then the pricing cards for two models of watches should also

have 3D models that it creates.

So this should just make a high end front end with a nice 3D modeling and

camera panning thing.

So here's our front end design test, the watch website in specific.

Okay.

Not quite.

I actually don't mind the font it's used and the watch co section.

It's just the models, unfortunately,

not up to par with what we've seen from other models.

Mainly the missing hands and things like that.

Maybe it's an LED watch if we were to give it the benefit of the doubt.

Okay.

So I'm a bit troubled, but

these watches are signific

antly cheaper than ones we've seen on other sites.

So maybe that's a reflection of the rendering here.

Let's give it some reinforcement.

No, I, I, I don't want that.

Stop it.

Okay.

So I responded alert expecting it to be like, I'm sorry.

Like, what do you mean?

It just went ahead and

thought that meant add an alert bar for the summer 26 drop.

So I had to stop it and say, sorry, I meant garbage alert.

And we'll see now if it tries to improve the result.

Got it.

You hit the garbage alert.

Let's take it out.

Nice.

The last one was too low poly and too much wood.

All right.

T-shot grade now.

I like it.

It understood intent.

So here's the second try of our watch website.

Okay.

It's still loading.

It just, it didn't make much of a difference.

It's slightly actually different.

It's just unfortunately not really.

Next up, I'm going to try a multimodal coding test just to see how it does.

I see the ZX Spectrum photos.

How dare you?

That is my own specific design.

Thank you very much.

So it's tasked with creating a 3D replica of this laptop that can be opened and explored.

We also need to have a functional keyboard where we can press things in the keyboard that it 3D models and

they will appear on screen.

So this is more of a test of its capability of seeing something and

then transposing it into 3D based off of a decent amount of reference photos.

All right.

Let's take a look at our photo to 3D model slash interactive website test.

Okay.

You know what?

That's, that's not bad.

It's not the best one I've seen, but

oftentimes models will invert the keyboard for some reason.

This hasn't done that.

Let's check the open close capability.

Okay.

This actually, it's not perfect, but

it's definitely better than some of

the open weight models I've seen do in this task.

Toggle RGB.

Toggle RGB.

Cool.

Reset cam.

Now, we should be able to type on the screen.

The problem is,

the screen is either hidden behind something or it's just not working.

But the key press events don't look bad.

I will say, those are decently solid effects there.

The close animation is not bad either.

Interestingly, no model really has gotten the hinge to be in the correct location,

which would be over here.

There's definitely some issues here, but lid open 160.

Oh, that's just frames per second, like, from this scene.

It's passable.

So, next up, I'm going to give this a 3D model test where it needs to create a 3D model of a V8 engine for my remote control car.

I have a 280 sized motor that I'd like to put inside of this engine model.

So, I wanted to make it something that looks realistic and

also is something that can be 3D printed.

We'll see how this does.

Interesting.

It just wanted to find out about the system.

And it's looking up motor specifications.

Good.

Using standard 280 specs.

So, my camera wasn't recording for the 3D model test,

which, you know, it does happen sometimes.

Basically, overall, we'll look at this once more.

It actually did a solid job of making a replication of a V8 engine.

Some of the things it included were actually very interesting and

levels of detail that I've not seen in a lot of other models when

doing this prompt.

Unfortunately, the big issue,

as we're going to see here if we open the SCAD file,

is that the hole for this engine is not deep enough to fit a 280 motor.

And even if it wasn't intended to go all the way in there,

the problem is this base here just cuts into the cylinder.

So, it would be physically impossible to install that in here.

I did also tell it to optimize this so

it could actually be 3D printed.

And unfortunately,

there's just a bit too much odd geometry everywhere.

And if we open this in a 3D slicer program,

we'll be able to see a bit better in reference to that.

So, this would be very difficult to print.

Just with, like, gaps like that, you would need...

I mean, it wouldn't be horrible to print, but

it would need a lot of support.

And then we have an issue like the air cleaner here on top of what would be the carb is just floating.

The valve covers are also kind of floating.

The connection's not great to the Vs.

But the things that it did do that were good that are just weird is the valve covers actually have,

like, individual hex bolts that show, like, them being bolted on.

And they're really quite well done.

Now that I look more at that, though, yeah.

I mean, that's cool.

And then the other thing it did,

it put an oil cap on one side of

the valve cover, which is kind of how it would be in real life.

So, just weird detail that it included like that that's not often seen in other models.

And then, unfortunately, it just wasn't altogether competently.

But it does show some level of promise.

And on first glance, just from, like, zoomed out,

it's like, whoa, this is really quite good.

So, that was my initial reaction that I failed to capture on camera because

I was not recording.

Nonetheless, we're going to move on now to a different task.

And that is going to be one that I unfortunately did not get.

I tried doing this in my GPT-5-6 test, and

it was the last thing we ran, and it just

froze up, and I was never able to play it.

So, I've just taken a screenshot of that prompt.

I had meta OCR it from this web chat interface, but

I will have it be built through open code.

I'm going to start it in plan mode here, because why not?

And we'll see how it differs, if it differs.

Got it.

Street eat.

You are BigFist McGee in a pastel low-poly city block.

Okay.

That name was not included in the prompt.

So, it came up with that on its own.

I did see some folks talk about how this could be,

like, funny to talk to.

We'll try that later.

Okay.

That was a good plan.

Open questions.

Now, the original Muse Spark, when I tested it,

it made a very fun Steve the PC Repairman offic

e game where you just go around and smash PCs and beat up NPCs.

So, it'll be interesting to see how this does.

Alright.

Let's check out Street Yeet.

So far, I like what I see.

One fist to yeet them all.

Hold mouse for a mega charge.

Look for golden objects.

That's a bonus Yeet.

Try to get the buses.

I dare you.

Again, we can punch pigeons, which was included in the GPT-5-6 game.

It's not included in the original prompt.

This is not bad.

Well, let's, uh...

Oh, yeah.

Alright.

The pedestrians are not quite right, but...

Hey.

Why can't I...

Okay.

Okay.

The cartoon style and stuff.

Alright.

It said try to yeet the bus, so let's just...

Let's just try this.

It didn't go as far...

Okay.

Now, that's actually...

And it has sound effects when we hit it.

Cool.

We're supposed to collect these, but I don't see how.

Let's just yeet this car.

Make it delete.

Alright.

I can't yeet the pedestrians.

Oh.

Oh.

Oh, I wish they looked more humanoid.

Is that a...

Pigeon?

I think it might have been.

Well, birds are meant to fly.

I can yeet everything here.

Elam, yeetus?

Why is it...

It's models.

It's just...

Okay.

The model's fun.

The Muse Spark.

Even the first one.

It's good at making fun or funny games.

Alright.

Let's try some multimodal coding.

Also, I'm gonna turn off this just rhythmic pounding noise.

It's questionable.

Yeah, I see it.

You're 10 yeets deep.

13,000 chaos.

14, 19 meters.

Far this nice.

I also see some issues.

Okay.

I wanted to include that the humanoid, like the NPCs,

are also garbage.

But nonetheless, we're gonna just let it ride.

Alright.

Let's check out our hypothetically improved Street E game.

I'm not really thinking.

Let's just try some creative writing.

So I'm giving it this photo,

which you may be familiar with if you watch the channel.

So thank you.

So I'm saying this is the cover for an upcoming psycho thriller romance novel.

Give me the title and back cover summary of the book.

And we'll see what it says.

Alright.

Alright.

Alright.

Title.

The way he holds me.

Let's zoom in a little.

Tagline on the front.

He'd do anything to keep her smiling.

Here's the back cover summary.

Everyone said Anna was still lucky.

When Mark asked her to leave the city and

move into his beautiful, crumbling Victorian home

three hours from anyone she knows, it was romantic.

When he deleted her social media because

you don't need anyone else, it was devoted.

When he holds her for photos, his smile is perfect.

Hers never is.

In the picture on their mantle, he is beaming.

His arm around her shoulder.

To their friends, to their neighbors, to the camera, it's love.

Anna knows better.

It's a warning.

Stay still.

Smile.

Don't pull away.

She has been planning to leave for months.

She has a bag packed in the attic and

a bus ticket hidden in a loose floorboard.

But Mark has been planning too.

He knows she's unhappy.

He knows she's been lying.

And he has no intention of letting his wife ruin their perfect love story.

A claustrophobic, slow burn,

psychological thriller about performative love control and

what happens when