Looks like the ZAI team is finally back
with a brand new model drop, the
official launch of the GLM 5.2, the new
open-source powerhouse. This is their
new flagship open model, and it is
looking like a frontier level
intelligence that is comparable to some
of the biggest proprietary models like
Pod Fable 5 in web design while being
significantly cheaper. GLM 5.2 2 is a
major improvement in coding agentic task
and it's built for long horizon
capabilities with a 1 million token
context window and it comes with two
reasoning levels. You have GLM 5.2 max
and GLM 5.2 high. Now the model is
released under the MIT license with open
weights available. But this is the first
time I've seen an open-source model that
is actually this good at web
development. It's not just a good at
open model. It's beating proprietary
models outright. It is actually
outperforming better than Gemini 3.1 Pro
and web design, which is actually nuts.
And from what I've seen, GLM 5.2 does
exceptionally well in front-end
development where it's currently ranked
number one on design arena, even
surpassing Fable 5 while being
drastically cheaper at around 1300 ELO
score. Like for example, just take a
look at the GLM 5.2 on the left and Opus
4.8 on the right. to build the exact
same landing page. You can see that both
of them do exceptionally well. And
honestly, it's hard to tell the
difference. But the crazy part is JLM
5.2 costs just 6 cents to generate this
version, whereas Opus 4.8 costed about
50. That's over six times cheaper while
being also faster and more token
efficient. Benchmark-wise, the GLM 5.2 2
is a new open-source king that is even
beating proprietary models like Gemini
3.1 Pro across almost every benchmark,
outright beating it. GLM 5.2 scores a
46.2 percentage on deep suite, and it
leads GLM 5.1 by a wide margin across
coding, use, reasoning, and general
knowledge. The model also scores a 74.4
percentage on Frontier Sway, just behind
Opus 4.8, which is wild. an open model
getting this close to Opus is honestly
just mind-blowing. Even on benchmarks
like Terminal Bench and Swaybench Pro,
GLM 5.2 is either beating all of the
proprietary models or coming close to
models like GPT 5.5 and sitting near 4.8
as in Opus. For GLM 5.2, Two, they also
strengthened the 5.1 context training
specifically for coding agents across
large scale implementation, automated
research, performance optimization, and
complex debugging. This results to long
context, which is where it's able to
perform quite well across broad scopes
and reliable in execution. There's a lot
of stuff in the AI space that I don't
really put on to the YouTube channel and
you can actually access it through my
free newsletter with the link in the
description below where you can
subscribe completely for free. On the
world of AI benchmark, you can see that
the JM 5.2 is quite high up. You can see
right now that it is ranked number fifth
and that is pretty high. You can see
that in most cases it does exceptionally
well with front end 3D as well as game
development. But there are a couple of
weaknesses with this model. It's not the
best in terms of debugging, reasoning,
and with its genta capabilities, still a
bit lackluster. It's also recommended to
use GLM 5.2 with high thinking. It is
going to get you the best results in
terms of cost to latency. Pricing wise,
the model is listed at $1.20 per 1
million input tokens and $410 for 1
million output tokens. Basically, the
exact same thing as the GLM 5.1 pricing.
To get started with this model, you can
use the world of AI vibe coding
benchmark as well as vibe coding
platform to start using the GLM 5.2 and
you can get started with this platform
for free. In my opinion, this is the
best way to evaluate the model with my
benchmark rules and methodology to
evaluate it on different sorts of
circumstances. Also seeing how
proficient it is in different domains
like front end all the way to backend
logic. You can easily also interact with
it through the ZAI chatbot. You can also
use it through the API and the open
weights are available for the JLM 5.2.
Now, to get started, we're going to be
testing this model out with a couple of
different front-end prompts. And this is
where we want to open this up in a full
page so we can visualize this a bit
better. But this is essentially where we
had requested it to create a beautiful
front end that has all of the different
packages that we had requested. And you
can see that with this output, you have
all of the scroll triggers that have
been attributed with this landing page.
You even have the background package
which showcases the shaders and then all
of the typographies have been displayed
pretty well with this generation here.
I'd requested it to create a soundboard
visualizer, an audio visualizer. And you
can see that this model does pretty good
with all of the visual elements. As you
can see, if I raise the volume, it does
great job in depicting how this actually
looks within the visualizer. You have
the ability to change the different
color themes, visual mode, which change
it to different elements. And you can
just see how beautiful this output is
with the visualizer. Even when you tell
it to clone different sorts of websites,
it actually does quite well in
extracting all of the elements from that
website and then creating a clone off of
it. And you can see that with this
Airbnb website where it did a great job
in creating all of the elements of a
landing page for Airbnb. You can even
click on the photo panel which displays
all of the images of the Airbnb which is
a pretty cool feature that most models
don't actually do. But guys, you'll
realize with the front-end taste of this
model, it's just truly different from
what we've seen from other models.
Everything from the scroll trigger all
the way to the different packages that
the model uses for a front-end output.
and it just does actually amazing with
all of these different elements even
when creating multiple pages and keeping
that cohesive paste throughout all of
the different pages. So that is one of
the best things about this model when it
is working on front-end development. It
also did a great job in displaying this
3D model of a watch and you can see that
with all of the different elements where
you can even tweak the different parts
within the watch and you can even get a
label that displays each of the features
within the watch. So you get a better
understanding of the mechanism itself.
Next is where I had requested it to
create a dungeon crawler game. And you
can see that this is an exceptional game
that it was able to generate. You have
the ability to maneuver through
different areas. And once you have found
a key, you can actually explore
different areas by providing the key.
And you can see I just unlocked that
door. You have different mobs. You have
different potions, items that you can
collect. And you can slay these
different mobs. And the fact that it put
this much detail into this output is
exceptional. And this is something I
didn't even see from the Opus model and
it's not even able to replicate the same
sort of quality. Let's now take a look
at the Mac OS clone. And this is where
it did an exceptional job with all of
the components. You can see that it did
a great job with the Finder app, which
looks quite accurate to how a real
Finder app looks. The top bar works. You
can actually close this window based off
of the top bar function. You even have
spotlight that has been added and the
menu bar on the top right which is nice.
Now if you are to take a look at a
couple of these different SVGs, it
hasn't been thoroughly generated
properly which is the only downside. You
have the Safari app but you have the
icon animation that thoroughly mimics
Mac OS. You have a terminal calculator.
You have a system settings which is
where you can change it to dark mode.
But there are a couple of I guess small
things that don't actually work properly
or haven't been generated yet. Now, that
is one of the downsides with this. You
can change the wallpaper, which is cool.
So, that is pretty nice. You can even
play around with a couple of these
games. You have a chess clone. You even
have Seduka game that was generated. And
then you have an AI assistant that was
also generated with this. So, overall, I
would give this an 8 out of 10. Next, ID
requested it to create a Spotify clone,
and it actually nailed it perfectly. You
can see that with the home screen as
well as the search. You can see that
each component has been thoroughly
generated, and it even plays music. Just
take a listen.
Next up is where it was requested to
create a Minecraft clone. I got to say,
this is one of the better ended
generations of a Minecraft clone because
the GLM 5.2 did a great job in creating
the inventory system. You have different
mobs. Looks like something is trying to
hit me, these different mobs. But you
have a lot of these functions that do
look like it was generated quite well,
but you can't really break or hit it as
the actual game. But regardless, it did
do a decent job. You have the ability to
place different blocks. There is a break
animation as well, which is nice. And if
you want, you can even find different
cave systems. So, if I have to dig, you
have full-on cave system that has been
added, which is insane. So, that is
really nice. This is a cave system that
I haven't seen with any model. So, that
is a really good sign. Now, I got to say
they absolutely killed it with the GLM
5.2 cuz this is a model that does
exceptionally well across all of the
domains that I showcased throughout
today's video, which is where on the
World of AI benchmark tool, it is ranked
currently as the fifth ranked model
across all of the Frontier models, which
is just incredible. and 3GS the model
does quite well as well because you can
see that with this generation where it
was requested to create an FPS shooter
and you can see that it generated all
the components you have the hue from the
shots being fired and you can see that
there's a break animation that was fully
coded out in 3GS. This solar system is
something that was also generated in 3GS
and I do love how it looks but it
doesn't seem to be functional based off
of the warp or like the speed of the
orbit. You can tweak that by managing
the time warp. And that is where you'll
see that it looks accurate. Our solar
system was accurately generated. Each of
the planets were thoroughly generated
with different attributes that represent
the planet. Like Venus, you have Earth,
which looks like it was generated
decently. Not the best, but it did a
decent job. Jupiter looks like it is
also quite well with this generation. It
even has the red eye. And you even have
the asteroid belt which looks quite well
generated with this output for the lava
lamp prompt which is something that
assesses how well the model is in SCG.
You can see that the dynamic movements
of the blobs is perfectly generated to
fit how a real lava lamp actually looks.
And that is one of the great things
about this model. It has good spatial as
well as physical understanding with its
generation quality. It adds in the
creativity while also making sure the
logic is there with every output. Even
have a small hue or an ambience that is
being outputed from the lava lamp. This
has to be the best procedural tree
growth generation that I've seen from
any model even over the Fable 5 cuz this
is where did a great job in representing
how tree would actually grow. Generates
it with ambience. to even have an
accurate description of the leaves and
it even showcases the output within the
shadow. So, that is actually pretty
unique of itself. Or you can consider
joining our private Discord where you
can access multiple subscriptions to
different AI tools for free on a monthly
basis, plus daily AI news and exclusive
content, plus a lot more. If you like
this video and would love to support the
channel, you can consider donating to my
channel through the super thanks option
below here. I'd requested to create a
Pelican riding a bike. And you can see
that it did a decent job in SVG. It even
out animated the wheels running. Now,
certain components don't look accurately
the best with this output, but still
regardless, it did a decent job. So,
overall, I do got to say this model is
incredible model. I'd rank it amongst
the top five models that are out there.
And that is quite an impressive stand
for GLM. And that's insane that ZAI was
able to do this with this open-source
model. So, it's really nice to see that
we have an open- source model being put
against these proprietary giants like
Opus as well as like GPTA. So, this is
definitely one of the best models that I
have seen in a long time and I should
100 100% recommend that you try it out
with links in the description below.
Remember, all of these tests that I had
generated were fully outputed with the
World of AI benchmark tool and Vibe
coding platform. So, if you're
interested, you can easily get started
with our platform with links in the
description below. You can also access
all these other links that I use in
today's video, which I'll leave in the
description. But with that thought,
guys, thank you guys so much for
watching. Be sure to go ahead and take a
look at the second channel, join the
newsletter, join the Discord, follow me
on Twitter, and lastly, make sure you
guys subscribe, turn on notification
bell, like this video, and please take a
look at our previous videos so that you
can stay upto date with the latest AI
news. But with that thought, guys, have
an amazing day. Spread positivity and
I'll see you guys fairly shortly.