shooting out.
I believe that was something I've seen .
The trees are quite reminiscent of a pool.
DeepSeek V4 Pro has just been released in general availability version
, which also happened with the V4 Flash model back about 12 or
13 days ago on 07/31.
Now, that model when tested proved to be a significant
increase in capability over the preview version that was previously
around .
And today we're going to be finding out if the same can be said
for the newly updated checkpoint general availability for V4 Pro
dubbed 0813.
So before we get into it , do feel free to subscribe so we can
hit 70K and then 100K with the plaque .
And with that , basically there's not a lot to say about this right
now because it seems to have been more or less silently released
.
There's no official announcement post yet on X ,
which I do believe is generally where we would see this .
But they do mention here just on the website a couple references
to it where if you use this model identifier right here for V4
Pro, you are being routed to the new V4 Pro 0813 version.
Additionally to that, it is also live on OpenRouter and this
makes me feel comfortable that it is actually genuinely released
and this is the model we're going to be using .
So , from the bit of information we have right here ,
aside from just the model identifier now being today's date ,
depending on what locale one is in, we can see that the model has
a million context window with a maximum output length of 384K ,
which really is a lot .
Now , something that can sometimes happen with DeepSeek ,
at least historically , is there may be a lot of thinking .
Now , this does support selectable thinking modes ,
as you can see if we do click on this .
So , it's possible to kind of tone that down ,
but in the interest of seeing the maximum potential performance
, all of our coding tasks for today are going to be run on this
maximum effort .
Now , I want to mention the following with a grain of salt because
I cannot definitively confirm that this is ground truth ,
but I would imagine it probably is .
There were some references and photos floating around which apparently
came from a group related to DeepSeek .
I believe they said it was a WeChat group about DeepSeek or something
.
And if this is to be believed , we can see that these would hypothetically
be the benchmark results for DeepSeek V4 Pro 0813 or the general
availability model .
And basically my main interest here is just seeing the leap in
capability over the V4 Pro preview to now because I was initially
disappointed with the V4 Pro Preview as well as the V4 Flash preview
.
They felt very undercooked and probably because they were named
preview .
That's actually like fair .
It makes sense .
But again , I want to reiterate the leap in capability between
flash and then the 0731 general availability version of flash was
absolutely massive .
This is like the best local model right now for like a dual 3090
setup .
So I mean in a niche , you know , but still if that leaping capability
increase from 12 .
increase from 12 .
8 to 62 .
7 would then hypothetically put this above things like Opus 3 .
8 8 and make it very very competitive especially that this
is an openweight model or will be .
That's also extremely exciting because this is not like horrifically
gigantic I guess when considering things like Kimi K3 or Qwen
3 .
2 .
It's still big at 1 .
6 trillion parameters .
It is a mixture of experts model with how many active ?
So if we do estimate that this is going to be the same size as
the preview version of V4 Pro .
It's a 1 .
6 trillion parameter mixture of experts model with 49 billion
active .
So it is still very very large but it's not as gigantic as like
closing in on a 3 trillion parameter model.
So additionally to that pricing and this is something that we need
to basically take just as it is right now because they mention
and have been for quite a bit of time that their API pricing is
going to be significantly increased soon .
Now significantly is relative to what it currently is which is
like very very very cheap .
So , right now , pricing per million input tokens for DeepSeek
V4 Pro is 0.43
35 .
So , okay .
No , I know what this is , but like last time I messed up on this
one and I said it was 2 cents and it wasn't and then I got roasted
.
So, this is 43 cents per million input tokens and 87 cents per
million output tokens is the current price .
Now , if this gets like a significant increase ,
say it goes up like five or 10 times , it would still hypothetically
be competitive .
So , it's all relative to what it currently costs .
And with that, I am going to be using this from within Cursor
.
When we tested V4 Flash , the general availability model on the
31st of July, I did also use it through Cursor because I believe
they had suggested to do so and the performance was very ,
very good .
I was happy with it .
So, I'm going to be using Cursor as well for this.
Now I should also mention that they did release in beta a desktop
app for Cursor or ChatGPT.
So I also have it routed and working through this desktop app as
well and it is doing a skate game for us in C++.
I would imagine that will take quite a bit of time .
So I did just want to start it and verify this app was actually
working properly on this system and it does seem to .
So keep in mind any mention of GPT or Cursor here is routed directly
to DeepSeek on max reasoning effort for the 0813 model .
And with that , let's take a look at our Okay ,
so apparently the OpenAI beta application set itself to be the
default for all file types .
That's infuriating .
All right .
All right .
Very Gen 3
aesthetic
I'm going to turn sound on because I did see some mentions of it
just talking about like sounds .
All right .
It starts in our notepad .
Can we resize this ?
All right .
Acceptable .
Minimize .
Unminimize .
Good .
The big one .
Is there a right click ?
Yeah , there is .
There is also the correct date and time in my local in the bottom
right .
And there is what I would imagine here to probably be the special
feature .
It is the holographic desktop feature .
So , we will save that for last .
Let's start out by looking at the files app , which I don't think
I've seen a files icon that looks like this before .
All right .
We have some nice transparency to the application backgrounds and
definitely a nice synth wave theme , which I do personally like
.
So , unfortunately , it seems our fake file system is just static
and we can't browse through it .
Nonetheless , can we create new things in it ?
Okay , we can .
Can we remove the new things we created ?
We also can .
Can we remove things that were there by default ?
Yeah , we can .
All right , cool .
That makes me a little happier .
And then up would just go there .
Oh , okay .
Not bad .
Decent bit of immersion , at least in terms of what directories
would exist were this a real computer .
Notepad we did kind of see because it opened with this just to
give us some information .
The only thing I like to see is will it save it as a text file
on the system .
No such directory .
Didn't I delete the documents directory ?
Hold on a second .
That's okay .
Oh , it remembered that I did that .
I might have messed that up .
So that's going to be that's on me .
Uh okay .
This right here is just not right .
It should not have a like visible text box .
That happens sometimes .
Help did give us things .
Let's just do neofetch .
That's actually pretty well done right there .
Neofetch or Neofetch .
Neofetch .
Still , that's that's actually all right .
I'm just being kind of mean .
Memory .
This is fine .
Oh , it actually put a bit of cleverness in there .
1 .
44 megs out of infinity .
This is fine .
Isn't that how much an old school floppy disc would hold ?
I don't know cuz those are older than me .
Next up , calculator .
Okay , so 47 47 47 8 8 8 376 .
I haven't done one of these in a while divided by six .
All right , settings .
Definitely seeing a bit of like Qwen 2.5
8B in here just in terms of some of the style here.
What is that and why do I like it ?
Aurora .
Oh , okay .
Sunset City .
This No one would understand that reference .
This looks like something one of my friends in high school was
making as a game project .
It actually looks like a ton like the splash screen for his project
.
We have customizable accents as well .
Difficulty easy , normal , and hard .
Oh , for games .
Okay , I probably set it to easy as you may be familiar with my
quick to anger game playing style .
Next up , let's try our games .
The speaker is on .
So , City Heist .
Oh , you absolute .
Okay .
So , you have to press enter to start and not click the button
.
Please shift gears .
Ah , okay .
Okay .
All right .
It's Oh , cool .
Oh , okay .
It's troubled, but the sounds for like the smack are kind
of funny .
Like , all right .
And we wrecked and we only grabbed a 100 red out of a thousand,
but it did work .
So , keep in mind we need to press enter to play these games and
not hit the button that says launch .
Oh .
Oh , all right .
Do I go through these ?
Yeah .
I I just assume based on the sounds .
I wonder if that's how like dogs are since they don't understand
language .
But like that to me seemed like a positive reinforcement sound
.
So , it makes me want to do it more .
And I think we did just discover something about how dog brains
work .
Okay , the score .
I don't know that we need that many decimal points just to show
a score .
I think maybe just rounding up to 734 or down would have been acceptable
, but I applaud it for its intricacy .
And then , oh , read me .
Okay , so our special feature is something like holographic .
And I would imagine having a couple apps open just when we trigger
that is probably the move .
So , oh yeah .
All right .
So , it just puts like an interesting card effect .
This is what it looks like .
This looks like an Apple Vision Pro like app demo .
This is the way they look from inside the Vision Pro headset .
Um , just if you have used one , you may resonate with what I'm
saying .
Okay .
Overall , it was fine .
Everything worked .
Nothing really blew me away .
But again , that may be a limitation of the test and less of a
limitation of the model .
So alas , huh , it's kind of neat .
We will continue .
So our skateboard game is still going .
And again, I'm using this with the newly released beta version
of the chat GPT desktop app .
I would imagine this will take quite a bit of time .
So let's check to see just like Oh no , cuz we already had looked
at the cost .
And keep in mind this is partially for the skate game as well as
the browser OS .
So 20 cents total across two separate tasks .
One which is still ongoing .
All right .
So , after 68 minutes and 24 seconds , Skate NYC is complete .
So , oh , you idiot .
You spoiled it with these screenshots and you don't even have vision
.
I'm infuriated .
Nonetheless , now that we know what the game looks like ,
I had to stop myself from saying something a little more colorful
.
Let's just take a peek at at the game .
All right , so here's our NYC skate game .
Should be in outputs , right ?
Yes .
All right .
It's not quite as good as the Qwen 2.5
model, but that model does have 800 billion more parameters
.
Does that say NYC skate on the ground ?
It's inverted , but it does .
And it says established 1999 .
All right .
Uhoh .
Sounded like the beginning of a Van Halen song
I don't know what exactly it is
All right .
Uh , let's just maybe not do that same thing .
There are sound effects , though .
I don't normally hear sound effects .
Can we actually move around?
All right .
Not quite .
I will say it's not wonderful .
Oh , that's actually kind of cool , though .
The way it like stylized the game title .
Nice .
Nice .
Well , we don't need to do the drum kit sim .
The trees are quite reminiscent of a poo, if I were to be honest
, but all right , the the legs are like kicking .
This may be the funniest skate game result we've ever seen ,
which is , you know , something .
All right , we have a street sign there .
The text is a little sideways , as are the building faces .
This is definitely a troubled result .
However , I do like the water fountain shooting out .
I believe that was something I've seen .
Okay .
All right .
I've seen that in, I think, the Qwen result.
Now , here's what I would like to just quickly do for some reference
.
This is the result we received for the C++ skateboard test from the
preview version of DeepSeek V4.
We can see right here .
I'm going to turn the speaker off .
This is pretty bad .
It was like so bad that I actually remember to go to this specific
thing just to showcase it .
So, the leap in capability based off of what we see right here
to what we just saw in the skate game that it had created for us
, which we can just look at in these screenshots here is still
pretty significant .
Next up , I'm giving it the Subway FPS test .
I've also added on to the end of the prompt .
Don't worry about checking over the script .
I will manually check it and if there are issues I will report
them back to you, just because this will really extend every
single result by like two times in terms of time-wise length
just in the interest of something I could do visually in like 5
seconds and then be like hey this doesn't work .
So we're going to do that just to maybe like keep things a little
tighter .
All right so it did finish the Subway FPS game .
The problem is it just keeps trying to do like troubleshooting
which it's not good at doing on this system for some reason .
All right , let's check out our subway FPS .
Main Rand Avenue.
All right , going to turn the speaker on .
So , it gave us two options .
We can either explore the station .
Okay, it's... Hey, you know what?
I think that is the best subway car that I've ever seen drawn in
this test .
And you're going to be like , well , it doesn't look that great
.
I know .
It's just that they never really include these .
And when they do , they're not really very good .
So this... actually, what was that?
If that thing starts moving I would have probably become more excited
than would be reasonable for you know responding to such thing
.
This is very clean .
It's simple .
It's nothing that's blowing me away to be completely honest with
you .
Queens Boulevard line .
Okay .
We have stairs .
We have some lighting elements and things of the sort .
We have exits if you want to go that side.
And then tunnels, like indicators for the train, a trash bin.
Floating particles more simple .
And we can't go up the stairs , which is fine .
We don't want to .
Now , of course , the other feature that was included here is the
FPS .
So, let's go back to the menu and then enter the survival FPS.
Let's see if it changes the lighting or anything .
Oh my .
That's a good weapon model .
Does it ?
How ?
Oh , it does .
I think .
Yeah , it does .
It's like smooth now .
I don't see any enemies .
Look at when we run .
It's like a waddle .
All right .
So , uh the big thing I'm noticing here is there are no zombie
enemies .
So , good reload .
The waddle is just kind of funny .
Weird .
Hey , look .
There are shell casings that actually drop .
That's kind of cool .
All right, this is like... yeah, it's kind of mid.
So , I'm going to give it a feedback piece .
Now , I stopped it because it was just like it had finished and
it was like I can't use this in the headless environment to check
.
And at that point , it's easier to just check it and tell it what's
wrong .
So , I'll give it that honest feedback .
So , supposedly our fixed and polished version of the Subway FPS
is completed .
Let's just take a peek at it .
The sound is on .
All right .
Interesting .
I think it swapped the positioning of the buttons .
So , all right .
Seven hostiles .
It's showing me in the top right .
I unfortunately just can't locate them .
I would have expected it to have done better .
And again , this is one where we could go back and reference the
previous model of this .
So , basically the preview version , it was significantly less
detailed than this was .
though unfortunately we don't have enemies and given the amount
of time it took as well as the kind of multiple times I told it
there were no enemies, I'm regrettably going to call this one a next up.
Now I'm giving this a difficult task to create a 3D model of an
RB26 inline 6 engine it must be something that can be 3D printed
with little to no support and fit inside of it a tiny DC motor
called an N20 motor because this is something that I've told it
should like work with my RC car project. Actually, I don't think
that's specifically in this prompt , but that's the whole point
of this, where it should understand and it will scale it
properly to that motor .
So far , I've run this test with a few intelligent models and I've
never really received anything that was actually worth printing
.
So , hopefully sooner than later that will change .
So , our 3D model task is pretty interesting because it's just
doing a lot of research on like the actual proper orientation of
specific items on this engine .
Like where is the downpipe , which is an exhaust component ?
What side is it ?
What is the intake side ?
What is the exhaust side ?
Intake plenum left or right ?
So , I do like to see the amount of like intricate research that's
going into the culmination of this model creation .
And it would be really awesome for like us to get a model for this
task that's actually printable .
It would be the first one .
So , in 1 hour and 10 minutes , we received our 3D model response
.
This is a very, very intricate bit of text about the thing
it made .
Now , of course , the other part of that is actually just taking
a look at what it looks like .
Though , once I remember what I named this folder ,
we will be able to take a peek .
Assembled display STL .
I would imagine that to be the case .
Please be good .
I've really, like, I need a good one for these.
Actually , it's sideways .
Son of a gun .
This did a really good job .
This is actually like so far the best one .
It's not perfect .
We have some sideways turbocharger issues , though .
I mean, this is all right. Let's peek at this in OpenSCAD
.
And there are a bunch of individual pieces for this .
So , we also want to see if the engine that we told it needs to
fit inside the little remote control car engine .
Let's see if it has any hope of properly fitting. Research—
It has a research folder right here just showing like a bunch of
images and stats for this motor .
Very cool though .
And here we have some better preview renders .
This is definitely well done .
Turbo side .
Okay .
Underside .
All right .
It's not something that's going to really easily be 3D printable
.
Although , I guess I can't make that determination because this
comes in multiple pieces .
So , it's not designed to be printed as we first saw it just in
one lump like this .
It's designed to be printed separately .
I would like to see .
Let's look now at just the single first piece ,
which is ideally where it would be placing the 3D printed motor
that would go in it .
Unfortunately , printing this without supports is just not an option
in any potential orientation that we see right here ,
which is a bit frustrating .
It almost seems like it has these posts and holes in it to be able
to snap fit on other parts, which is cool.
Okay .
Yep .
And I think those are definitely like attachment posts ,
which is pretty cool .
This one would be able to be printed more or less without any supports
if we were to put it flat .
Then this should have the Nissan logo, or at least the text "Nissan"
written on it .
Yep .
And this one's quite well done .
No issues here .
The oil pan .
Oh , no .
Okay .
This is where we decide we're not going to be
trying to print this .
I'm not disappointed , but I'm not super impressed either .
So , I suppose I am lukewarm when it comes to judging this .
Definitely , if we just look at this as a render ,
even forget about looking at the individual parts there ,
cuz it's not really going to be feasible to print this ,
it did a really decent job just assembling it .
The research that went into this, I think, is probably
like I hate to sound like an AI here , but like the story isn't
the model , it's the research that went into it .
That's like generic AI slop.
It really did a lot of decent research here .
I mean , did it find this and download it just to see ?
No .
Okay .
So , this is something it maybe made or maybe probably found online
just to be able to better assess what the size of this asset is
, but it took a lot of care to ensure that what it tried to generate
would have hypothetically fit the specific specifications that
were denoted in the original prompt .
So , not half bad .
I'm also going to give this a simple front end design test where
it needs to create the watch website .
So , it has to have a pretty cinematic panning effect over a 3D
watch asset that it creates .
I did specifically denote that
the watch needs to be a 3D model .
So , we'll see how competent it is at putting the entirety of this
together .
So , in 1 hour and 30 minutes , we got our watch front end result
.
It again seems to have a pretty substantial like just list of what
was built here .
So , let's take a look and see how it is .
a beautiful watch front-end website.
That's actually, it's so close to excellent.
The big problem is this , I would assume , is the watch strap sticking
out .
This is 3D, so don't be fooled by its 2D-ish look.
It is actually 3D .
It's slowly panning , too .
I really... Okay, so a couple things I do like: the wooden background
and then this like leather card or pad that this watch is placed
on .
In addition to that, the face looks really good.
The materials look very good .
It's just like, overall, it's a high-end aesthetic,
which is exactly what we go for when performing this test .
Now , that's actually that right there looks good .
The... Let's just sit and appreciate it for a minute.
Okay .
Yeah , that maybe I could have done without that ,
but it's definitely quite something .
And then look at the way that goes away as soon as we start to
scroll down .
"Quiet luxury measured in seconds."
"Sunburst dial."
Okay , it has like a material card here .
The sapphire crystal is perhaps something I would have maybe omitted
from this section , but the dial looks good .
"Two watches, one point of view."
Please render them .
Oh , look at that .
That is sick .
That really looks pretty good .
I love the way they're placed on the table .
Really the big thing letting it down is just this leather slab
is inhibiting our view of these beautiful straps .
This one is the stainless steel option .
This one is the gold option and priced accordingly I would imagine
.
Yeah .
Although I would say the jump in price is like a negligible
amount to get the gold one .
I'd probably go with this one personally .
Oh , cool .
And these ones we can freely rotate so that helps a bit .
There is no movement to the watch hands but that's okay .
I'll forgive it for that .
But these are pretty darn cool .
These are well drawn , I would say .
Especially the numeral indicators .
Look at the way that they actually have multi material on them
.
So , they're not just like one static material .
It's gold and then black infill or silver and then black infill
as well .
So , that's some attention to detail I do like seeing .
As well as the kind of brushed material look on the stainless steel
and the gold .
It's probably weird we're like assessing like a 3D watch ,
but it just shows like some capability .
Okay , that looks pretty good right there .
And then we have a clean footer as well .
I like this .
I'm happy with this .
Even though it's like not perfect , it's darn close .
So, next up, just here from within Codex.
And I can't show this full prompt because it has an API key .
I'm giving this the full on cinematic Steve the PC repair game
test where it needs to create a cinematic game .
It needs to create voice assets for these characters just by using
an open AI API key that I've given it and it needs to make a fun
game that is enjoyable and playable .
I will assess this is probably going to take north of an hour .
So, we're just going to basically leave Codex here in the background
the beta desktop version of it to complete this task, and we'll
maybe work on some other things in the meantime .
So , our Steve the PC repairman cinematic game completed in basically
65 minutes .
So , a little over an hour .
All right , here's how to play .
I'll probably just do what it says here , not assume that it's
already running this in some server for us .
So, okay, episode one, the Azure Palm job.
Okay , click to begin .
Headphones recommended .
We have a speaker .
All right .
My name is Steve .
This is good .
I fix computers .
Most of my clients are old enough to remember when a hard drive
was a luxury item .
They pay what they can afford and I make it work .
Mostly honest business .
Oh , thank you , dear .
The screen kept telling me the date was 1980 and I thought, all
fixed.
I have been traveling long enough .
Fresh battery .
Your computer now believes I will bring you oatmeal cookies
on Tuesday .
Same as the rest .
Give Mr.
Buttons a scratch behind his ears for me.
Let me walk you out , ma'am .
I'll be right with you , Mr .
Thomas , the speech overlays itself .
A good boy , Steven .
If that quiet man needs fixing, you be patient.
Punctual .
I like punctual .
Door is unlocked .
Coffee is warm .
A man .
What is the job ?
Came for 2 years.
My life .
See the the head tilt movement back and forth is actually there's
three people talking now .
And he forgets the entire name. Charles buys him a penthouse private
security, security wants his legend. Some people would like to ask
him questions. Him slowly. The client pays 100 upfront. I would like
it's also up in the case. All expenses, first class, two tickets. This
is why we work together. Together, the top tray is a normal file
on a laptop. It's quite my second favorite thing.
Professional .
I will not ask what the first is .
The phone rings when the helicopter is ready .
40 .
Cut scene over .
Thank goodness .
Click to take control .
All right .
Aside from the sounds being just totally messed up ,
this is actually pretty decent .
Okay .
The cat is... There's definitely some geometric issues.
What are we supposed to do ?
Take the briefcase when you're ready .
Okay .
Where is it ?
Well, press buttons are like, "Don't look at me like that."
Buttons .
Easter egg .
A man's got to eat .
Okay .
Where's the briefcase ?
All right .
We have inventory .
Zip ties .
Okay .
E to continue .
All right .
Is it going to put us in Miami?
Nice .
Palm trees , humidity , and one very bad man ,
Mr .
Colton .
Champagne perhaps .
I'm just going to click through cuz it's like like black like my
well, coffee asset at the resort.
Poolside at noon .
Oh , poolside .
17th floor enters.
Camera coverage .
Elevator .
There is no dog .
Behind is the west service stairwell.
Well , I think we just saw a glimpse of the pool .
All right .
Oh , great .
Azure Palm Resort .
Click to take control .
It's even got like the like low rent wallpaper that you'd expect
at one of these joints .
Oh , wow .
This is actually like kind of intricate .
Oh , good .
Okay .
I was worried we wouldn't be able to just walk through the walls
, but okay .
We can toss a coin in the fountain .
Your wish .
No more syntax errors.
50 50 .
It seems like Steve's shop and the airplane scene is like oddly
placed inside of the location we're supposed to be at .
Let's just ignore that .
So , check in at the front desk and then reach the 17th floor .
Welcome to the Azure Palm .
Checking in , Mr .
Me.
Coulter.
Steven Coulter .
Mr .
I believe I have a reservation .
Room 1408 , partial ocean view .
I do apologize .
The ice machine on your floor is out of order .
Story of my life .
All right .
We now have a room key in our inventory .
So , is this the elevator ?
Oh, is that right?
Okay .
West service stairwell .
Let's just assume this is north .
Good. Like, what?
Well , that is frustrating because it gave us this really like
anticipatory music .
I'm sure there's a far better terminology for what I was about
to say .
My name is Steve .
Oh , no .
Don't .
Most of my clients are old enough to remember when a hard drive
was a luxury.
All right .
We we got most of the way .
I think this is a test that more often than not ,
we don't complete it .
I will say the elements of this, judged individually,
Definitely the pieces were all there .
I don't know that the puzzle made the correct picture that was
shown on like the puzzle box , but it was pretty darn close and
I saw some—truthfully, this did a better job than I expected,
especially considering like the subway scene game and things like
that .
So , I'm actually I'm okay with this .
So , the final thing was of course the street yeet game ,
which is weird because this piece of text moving up and down like
that while this one remains completely static makes me uncomfortable
in some level .
So, this is a Bijan issue.
All right , good .
So, I neglected to, um, just run it in a server.
One fist, zero mercy.
I already like this .
Oh , it seems like Good music .
All right .
All right .
All right .
Let's just explore the city real quick .
I don't even know .
Okay, so I saw a car get—like, this one's actually kind of fun
.
Excuse me .
Excuse me .
Pardon me .
It does like a... like a... Okay.
I don't know if we hit...
Is that a Can we hit the bike ?
Bonk .
Uh oh .
Jay walking pedestrian .
Can we punch this parking meter ?
Yeah, we can just hit everything except the buildings.
All right .
There's unfortunately the orientation of some of these things is
not quite right .
We can see that we're moving sideways when we move forward .
And the cars are also Sorry , pal .
Oh , and it's interesting how they uh interact with the scenery
as well .
Can we eat this statue ?
I don't think we can , but Oh , yeah , we can .
All right .
I will say the cartoonishness of this game is actually kind of well
done .
I don't know how easy that is to assess just watching someone play
it , but like the pacing is decent .
I think maybe that's attributed to the music as well ,
but the way that—when you actually... All right,
the loop is getting a little repetitive, but when you actually yeet
the things and they also go and interact with other items in the
environment .
So like how that will bounce off the building and then maybe hit
other things in the path .
That's pretty cool .
And this was actually quite arcade style and playable .
Our score is going up .
We have combo multipliers overall .
38 here is a success .
All right , so that is probably going to conclude our first look
and test of the general availability DeepSeek V4 Pro ,
the 0813 model designation .
Now , this is definitely an improvement over the preview version
that felt pretty undercooked and I was not at all impressed with
it .
But if we look back at some of the tests that were run in that
video , which is on the channel , we ran some of the same ones
today .
This is a significant improvement .
I would say though , I think the flash model impressed me more
than this , especially considering the size of it .
This has a bit more competition , but again , it is a big leap
on its predecessor , which I guess is what we like to see .
I was impressed a few times with some things like basically results
within results .
The cinematic Steve the PC repair game , it actually had some elements
to it that were like , hey , you know , it kind of thought this
through .
Just the way when they were speaking , their heads were like moving
back and forth .
Not the weird chin thing that was happening .
I'm not sure what happened there .
The music it included in both that game and then the Street Yeet
game were actually like they definitely fit the scene that it was
drawing here and it did a good job making the Street Yeet game very
cartoonish and the playability of it felt good ,
I guess , for lack of a better term .
The model that we got of the engine was actually it wasn't printable
.
So , unfortunately , maybe it failed to like check that box in
the specific task .
However , if we just look at one of the previews right here ,
it did a pretty decent job .
Definitely better than the other models I've tested this with .
Though , I have not tested this specific prompt with any big beefy
Frontier models yet .
So , keep that in mind as well .
But, it definitely has some chops here.
And the amount of research it did to actually get to that point
was quite a lot .
We have all the things in here downloading these images when it
doesn't have vision .
So , I'm not quite sure what to make of that .
But nonetheless , I think probably the biggest point of frustration
for me is it just goes all out when trying to actually ensure that
every aspect of every line of code it made is correct .
And with a model like this where it does not have vision ,
sometimes it takes longer and there is a cost thing to it as well
.
Assuming you're going to be paying to use these models ,
it just sometimes is simpler to have it make the result ,
not try to do like overviews and checks and follow ups .
you look at it yourself and then tell it what's wrong and tell
it to fix it .
Having to just go all out and it trying to fix it adds a lot of
time and expense and tokens to the task , which maybe is more of
but I find that's something I've consistently had frustrations
but I find that's something I've consistently had frustrations
with as well as with the flash model where if I'm having them make
like a say street yeet, it's easier if they just spit the script
out , I test it and say , Hey , the buildings are disappeared .
and then it fixes it rather than it trying to go through and like
well let me try to spin up a headless browser and then I'll take
a screenshot and like do a heat map because it can't actually see
it or some other function .
So it's a little token-happy.
I did run it on the highest potential reasoning mode .
So that could be attributed to it as well , but I still would like
to be able to do that without it like going all out to try to make
sure it's polishing every aspect of the script because I want max
performance .
I just don't need that level of depth .
So , that may be a personal complaint , but you know .
So , for the final thing , let's take a look at everything we spent
for the entirety of this test .
We can see right here that all cost 1 .
91 .
So , that was pretty darn cheap .
And you do have to remember this is going to be increased, their
API prices .
They say significant increase expected , but I mean a significant
increase might be like , okay , the entirety of this test was 8
or something , which is still in perspective not horrible .
The Qwen 2
8B max test video that we did a few weeks ago or like a week or
two ago that was about 32 .
So keep that in mind as well .
That may be a bit more performant than this , but there is going
to be a pricing consideration as well .
And this was very very very cheap .
It won't remain that cheap , but even if it gets five times as
expensive , it's still not horrible .
I suppose that's going to conclude our first look and test of the
0813 Deepseek V4 Pro model , the general availability version .
And this is now out of preview .
And that's just pretty exciting .
Basically, now I think I'm going to test Grok 3
because that also came out today.
And then I believe Friday Qwen 2
72B is coming out.
So that'll be exciting as well .
Seems like there's been this has been like a pretty busy week in
terms of model releases .
So that's cool also .
And when MuseTalk 1
2 comes out , that's going to be awesome .
I don't know how big it is , but I'm very excited to try to run
it locally .
So with that , if you have any questions , please feel free to
leave them in the comments .
And thanks for watching .