That looks like a night time, like party room. Uh not that I would know. That's a strong word. Strong words get people hurt, friend. Oh, is this
Stevie hard talking? Today, we're going to be taking a look at the newly released Mistral 3.5 medium 128B. Now, this is a really interesting model
release because this is a dense 128 billion parameter model, and all of the recent models around that 120B
size point have been MOE. So, things like Google's Gemma 4, I think it was a 122B, or maybe that was a Qwen 3.5 122B, but even GPT-OSS 120B, all of
those are mixture of experts model, and this is not. This is a dense model. So, there are pros and cons to that. Pros being that hypothetically, it
should pack a deeper level of performance than an alternative MOE. However, the cons will be that it
requires a significantly larger amount of compute to run at a reasonable speed on a device locally, as opposed to an MOE model. So, this is a pretty
unique release for some other reasons as well, and we're going to begin just by taking a look at some of the highlights for this model, and then we'll
be running it both locally for at least some of the tests because it does run a bit slow on my
system, and we'll also use it just through the API in open code for some agent decoding tasks. So, before we get into it, please do feel free to
subscribe as I want that 100K plaque, and we are now closer to 100K than we are to zero, which is quite exciting. And let's just take a peek now at
some of the interesting things about this model. I think the first thing I'd like to bring up that is
interesting, aside from it being dense, which we've already touched upon a bit, is that this is actually multimodal. So, this has the ability to take
images in and then respond to them in text, but it's cool to have a big model like this that can natively see images, and that allows us to do some
tests where we give it a photo of a UI and then actually have it generate the UI in code, which is
always fun to do to test vision and programming capabilities. Now, interestingly, we see that this model is replacing multiple previous models, so it
is both replacing its predecessor Mistral medium 3.1 and DevSkrol 2. So, hypothetically that tells us this should have some pretty decent coding chops
if it can replace both a more coding focused model and a general purpose model. Additionally to
that, this does have toggleable reasoning, so you can run it with no reasoning enabled, which if you're running it locally may save you quite a bit of
time, but you can also have thinking enabled as well. Additionally to that though, I'm not going to be running this today. They do have this thing
where they say it is an Eagle model, and if we click on this, this is essentially a model that allows
you to perform speculative decoding with the Mistral 3.5 128B model, which is cool to have because in a big dense model like this, anything that can
speed it up is very, very cool, and that is basically exactly what that Eagle model is for. Now, I am a bit delayed in testing this because as we see
in this little red warning box, there was an initial issue with some of the GGUFs. There was a
looping issue or something of the sort, which made it so I didn't really want to test this model until things were properly fixed because I do want to
run it partially locally. So, I realized I did have the wrong Mistral Medium blog post open here. I had the one open for Mistral 3, which was like
over a year old, so I became quite confused when I looked at the benchmark and saw it was being
compared to Sonnet 3.7. This is the real benchmark JPEG that they have, and we see here this is being compared to Sonnet 4.5 and hypothetically is a
competitor to that model in pretty much every benchmark we see right here aside from this banking specific one. So, I'm very excited to get to testing
this. To begin, I would like to initially use this locally. I have this loaded in on my Mac Studio
M3 Ultra 256 gig system. This is the Q4KM quantization. I have ensured that thinking is on, and all of these sampling parameters are correctly set for
this model to be running in this specific framework or task, I suppose you could say. So, of course, we're just going to begin with the browser OS
test V2.5. Now, we're going to notice the thinking tags are going to appear, but they're not getting
parsed in an aesthetic manner, so we're not going to see the traditional like reasoning window that would appear right now. We just see an opening
think tag and then we're going to see a closing thing tag. Now, I find it interesting because I also have tested the Q8 quantization just to see speed
and it was running at around 5.3 tokens per second on the system for the Q8 quantization. The Q4KM is
not insanely faster than the Q8 is. So, I'm just kind of curious and something interesting to see. I do want to just test this with the Q4KM because
in the interest of time and also we will be using this predominantly through the API as connected through Open Router. Um it's quicker to just do the
Q4KM. Now, our browser OS is still being built and as a matter of fact, it is actually still
contained in all of the thinking capability here. So, this hasn't actually started generating the final response. Being that is the case, I'm going to
simultaneously just through Open Code here run another test to see how it performs agentically. So, right here what we're going to be doing is the
self-contained C++ skateboard test. I have put this in its own directory and I am just initiating this
from within Pine mode and we can see here. This is just through Open Router where they are serving Mistral. I should make a note that for API, the
pricing in this I believe was a $1.50 input per million and $7.50 output per million, which is quite pricey for the output. So, I'm very interested to
see how it performs here and I'm more inclined to judge it harshly as running through the API because
of the associated costs being fairly steep. So, as I was a little curious just because the speed was really nice here from Open Router, it is running
at 81 tokens per second just being served through Mistral. So, this is what's powering our Open Code instance right now, which is good as that's
definitely a proper speed. So, we did get a successful write of the self-contained skateboard game. There
were some errors during compilation and it was also doing some searches to find the specific dependencies and things that the system has to allow this
script to run. So, it's identified these issues and I believe it did outline kind of a small plan in order to fix these. We can see there was one
specific tool issue that we had right here where it was trying to write something. Additionally to
that, um if I scroll up right here, we can see that it was planning out to fix these three specific issues. So, that's what it's currently doing. All
right, that actually looks really promising. So, I was just sitting there watching it and it did a test run, obviously, some form of smoke test just
to make sure it compiled and worked. All right, as we check our skate game. Okay, I'm going to You
know what? This is Oh, so we must we're inevitably on like a rail or something right here. Space is to ollie, arrows and air is to do tricks. This is
potentially good because like look at the actual water effects. I do see some pedestrians and things there. The big issue is something's just not 100%
right. Fortunately, this model is in fact multimodal, so I think I'm going to provide this with a
photo of basically what we're seeing here and that should allow it to reference the issue. So, I've told it does run, but there are some pretty severe
graphical issues and I have included a couple of photos in the directory for you to be able to see and remedy these issues. All right, it's saying it
can't view images directly. Um I don't know what to say about that. So, I'm doing the best I can to
just describe the issue, which is basically everything's messed up. So, it's been working on trying to fix the skate game for quite a while now. We
can see our context window is basically 50% used up. So, it's showing 140k right now, which is fairly lengthy. So, it'll be interesting to see how it
is actually able to fix these issues over a longer context length than we would be at where this just
a zero-shot game test. I don't know 100% what this is going to do, but we see that it did finally try to compile it and there were no errors. Okay,
good. So, unfortunately, we've received I don't know that I'd say that's a worse output, but because it does seem like we It's tough. What are those
spinning green things over there? I can't really get to them. Okay, if I go Is that a trash can? All
right, well, perhaps the C++ uh the boardwalk aesthetic may have thrown it for a loop. If I adjust on the older style skate game test where it's just
a simple skate park, we likely would have had a better result. As I do see like some slight movements. There's feet, there's the board, but overall,
unfortunately, I'm going to say this is not really a wonderful result. So, partially due to some
impatience, I am also just giving it the browser OS test here through the web chat interface from within Open Router, and we can see here the
reasoning is happening at a significantly higher pace. So, I'm inclined to still let this go for LM Studio while it is working right here, and we can
see it did generate the entirety of the script that is hypothesizing it is going to create, but because this
is still all contained within the thinking tags, we're basically going to have to have that regurgitated in its entirety now though. Although, I'm
going to be honest with you, I have half a mind to just copy paste the initial thought process result and see if it actually works. All right, and
that was 874 lines, so that's definitely a fairly reasonable result to assume that we may actually have
something working here. Let's just see if we did get anything. Okay, you know what? We did. So, this was all, remember, contained from within the
chain of thought, but it basically generated the entire script prior to finalizing its thought process. So, more or less, this is what we would have
received had we waited. We don't have a right click, but we do have a correct time in the locale at the
bottom right here, and we also have a simple background for now and a task bar. Let's start with our start menu. Okay, and we just have a simple list
of all of the applications here. Now, I am happy to report the first app cuz we like going through them sequentially is the GTA clone. Oh, I opened it
twice by accident. You know, this is so odd because if this was a better I I what I'm trying to say
is the app window is way, way too small, but it actually did do this to a sufficient degree I suppose could be said. The buildings look good. There is
a basic car model. There's really nothing else going on, but in some ways this would check some of the boxes for GTA clone. The steering and movement
of the vehicles actually kind of interesting. So, nonetheless, I'll close that. We'll just try our
next game, which was a space shooter, and unfortunately, we don't really get anything there. Terminal, okay. We can open or change wallpaper. I should
probably try that, but there's a specific app for that, so we'll wait. Let's do open GTA. Okay, and that does work, so that's nice to see. So, we do
have a somewhat functional terminal. Next up, we have our text editor. Simple, but most are. We have
our calculator. Everything is laid out sufficiently. 58 * 6. I've done this one before, 348. Finally, then we have our wallpaper. Ooh, okay. Now, this
is again, it was we pulled this from the chain of thought, so we can't definitively say this was its full result, but more or less, it was probably
pretty close to what we would get. And we can see right here the think tank had just ended to the
point where it's generating the script now. And in the meantime, I am going to hop back here to where we ran the same thing from within Open Router,
being that it has finished, and I just want to do a comparison because we have to keep in mind this is a Q4KM running locally. So, let's see what the
Open Router version being served from Mistral would have done differently for this browser OS. And
here is our Open Router browser OS. You know, that's actually kind of interesting is these are really like quite, quite similar. Interestingly enough,
the wallpaper seems to be broken for both. Okay, this has a a prettier UI to it in terms of its file explorer and things of the sort. We do almost
have a Windows XP style. I'm going to initially try to see if there's a way to change the wallpaper to
actually have something that loads. Now, it is very possible this is using Unsplash image links that perhaps are not live currently or something else
maybe awry, but I have to say I actually feel better about the local one we ran prior to this because more or less these are similar. I'm obviously
very interested to see the GTA clone. Okay. Overall, really a very, very similar result between the
two, which actually speaks well to the performance of the local Q4 KM that we have running because everything here is definitely related in some way
or another. All right. Next up, we have a 3D maze game. Now, I do see something. Okay, I see a little green thing in a what appears to be a maze. I
think the big issue is that we can't actually change the view to properly be able to see what's going
on there. So, overall, it's interesting. These are basically sufficient. In some ways, this one was almost better, which is curious. So, I'm giving
this a website design reference photo that was generated with, I believe, Nano banana if I can full screen this. I've told it to just basically create
a website from this reference photo, do a really good job, and make it very high-end. So, we'll see
what it does. It is still currently reasoning, and this will be a cool test of its vision capabilities coupled with perhaps some of its front-end
design aesthetic capabilities. All right, we have received our mock-up from the reference photo for this website. Okay, unfortunately, the image links
here are broken, which definitely is going to let a result like this down regardless of what happens.
And I do actually believe what it should have tried to put here or what is not appearing there were the actual reference photos for the sample
companies that were included in this mock-up. So, what we should be seeing there that is unfortunately not loading is Google, IBM, Microsoft,
Salesforce, Tesla, etc. We do have the top effect here with get started. If we scroll down, we can see, okay, there
are some hover effects and things of the sort here. The Nexus Sphere platform in action. This is good to see because it did actually emulate some of
the charts from the reference image. Additionally to that, we do also have the sample like customer satisfaction testimonial things and then a simple
footer and everything like that. Okay, I'll say it captured some of the aesthetics of this reference
result, but not necessarily all of them. Um though, you know, this is a fairly difficult task because this is just not one simple image of this
website. It is three separate images that it has to then put together to create one single page. Next up, I am giving it the beautiful static subway
scene just through open router here out of pure curiosity if nothing more. I am actually still allowing
this from within LM Studio to finish out the complete result for the initial browser OS test we had initiated because one, I do want to see how this
differs from the preemptive thinking one that we copy pasted. So, here's our subway scene result is just the first one where it has to create a
beautiful static subway scene. Okay, sometimes when these load in, it like puts us outside a wall or
something and causes this Unfortunately, I don't know that's the case here. Let's take a look in the developer tool console and see. Okay, we do have
some specific errors. So, I'm giving it the errors. One of these is a nothing error because we're just not running it from within like a Python
server, but additional web server. Additionally, there was a syntax error in one of the later lines, so it
will have to fix that as well. All right, I've replaced the code with the fixed code. So, hypothetically, if we refresh this, we should get a result.
You know what? Uh-oh. Why can I not WASD move? This is actually decently aesthetic result, I think at least from Oh, okay. We could move. I think it's
just Let's check our brightness slider. Okay, very good. That does work and I would say the
emissive properties of the background images are pretty cool or background assets, whatever you'd like to call those. That looks like a night time
like party room. Not that I would know, but we have So, I'm noticing Let me just try to refresh into this. Okay, so it has like we should I noticed it
was almost delayed like when I regained control of the pointer by pressing escape, then it would move.
You know what? This is going to be a pretty good open code test, I would say, to improve this agentically. So, I've initiated this from within plan
mode, just telling it that it should be a little more detailed, and the WASD movement schema is not really working 100% properly. So, I'm going to
start it from within plan mode, and we'll have it initially fix these basic results. And then whatever it
ends up creating following that, I am going to then tell it to make it an FPS game. All right, so we had a few additional errors when trying to run
the fixed subway scene, and it basically said the file had become corrupted because of editing Oh, no, you All right, unfortunately uh this is a
unusable result. Basically, it became significantly worse than this to the point where I don't even now
want to try to turn this into an FPS because I don't know if it's actually going to successfully be able to do so, to be completely honest. All right,
I'm very happy to report that basically from the beginning, we've been running this browser OS test with the Q4KM Quant. As you're aware, it is now
close to finishing up, so we'll be able to see how many tokens this was overall, and the token speed
as well. All right, that was a total of 31 Basically, 32,000 tokens at a speed of 7.23 tokens per second. Now, seven tokens per second is it's
passable, but as we saw here, I I genuinely think this took like close to, if not over, an hour. So, here's our full completed browser OS with the Q4
local Quant, and so far I do have to say this is appearing to be the best one we've received just based off
the fact that it has this very interesting um tube style background. So, of course, we have the clock in the bottom right, which is the correct time
in this locale. There is no right click, which is fine because it's not specifically denoted that it should have one. Our start menu is definitely a
bit more aesthetically pleasing than the other results we've seen, and this is curious because we ran
this through open router where the model is being served by Mistral, and that result was so far significantly less nice than this one. So, this Q4 is
stacking up nicely locally. Okay, so perhaps unfortunately, we're going to start to hit some roadblocks here. Oh, I opened it twice, so Okay, but
functionality here has definitely been traded in lieu of aesthetics. All right, so still I think the
background just kind of had me falsely believing this would be a little better, but perhaps not. So, this model is multimodal, and I have heard over
the years that a lot of folks do like using the Mistral model specifically for creative writing. So, we're going to combine these two tests. I have
disabled reasoning for the respect of time, and as well as I don't necessarily know that it will make
as big of a difference whether or not it has reasoning when doing more multimodal and creative writing tests. I have given it this historic photo. If
you do watch the channel, thank you, and you are familiar with this. Telling it this is the title for an upcoming book. Generate the title as well as
a chapter outline. Now, we're going to notice that this is the same model exactly that we used for
the browser OS. It is just with reasoning disabled, which is something you can do in the actual like prompt template here from LM Studio for this
model. It's going to take significantly longer for prompt processing because it's not only ingesting just the simple sentence, but it actually has to
look at this photo and get information from it and things of that sort. So, as we see here, there is no
thinking, and we immediately begin to get the result. I'm going to try not to look at it until it is all finished. Actually, I may regret that at
seven tokens per second, so we'll start taking a peek. All right, so our book title is Love Beyond Measure, A Journey of Heart, Resilience, and
Unbreakable Bonds. Chapter Outline. Part 1, Foundations of Love. Chapter 1, First Glances, Lasting
Impressions. The moment they met that the sparks flew. Early challenges and how they overcame them. Chapter 2, Building a Life Together. Shared
dreams, compromises, and the early years of marriage You know, already this is kind of like basic, so I'm going to tell it that. So, I've given it the
feedback that this is pretty generic. Give us something spicy and fun and interesting, okay? Thick love.
Thick love, okay. Another time that this specific photo has seemed to instigate some form of weight-related remarks from any model that gets it. An
unholds barred memoir of passion, chaos, and unapologetic devotion. Already we're off to a better start. I do just want to say Chapter 1: When the
earth moved and not in a good way. Again, perhaps subtle tones of perhaps like weight mention. The first
time they met was it love at first sight or a disaster waiting to happen. It's given us a spoiler, it was both. The unfiltered, messy, real first
impressions. Someone may or may not have spilled a drink. Chapter 2: Love, lust, and the art of not giving a damn. Okay, very good. This is definitely
more creative writing Mistral style. The early days of raw, unfiltered affection, late-night talks,
impulsive decisions, and the kind of chemistry that makes the rooms feel smaller. The first fight and the makeup that followed. Details may or may not
involve broken furniture. That sounds disturbing like not good. Chapter 3: Against all odds, including our own BS. Oh, wow. Okay, so I'm going to have
to blur this out because it did include a swear and that's good, so the people who said it
wouldn't last, the doubts, the insecurities, and the sheer stubbornness that kept them together. How they turned "We're doomed" into "We're
legendary." Okay, I'm not going to read through the entirety of the rest of this, but I am going to say just based off of the feedback I gave it here,
it completely overhauled and changed it from the kind of cookie-cutter generic overview that we would
historically get to something that is genuinely filled with creative writing as well as some creative um lingo as well. So, this is definitely much
better. So, the next thing I want to do kind of based on that is give it a role-play test. Role-play as Steve Stevie, the PC repairman. And we'll see
what it does. And we'll we'll follow up with it and Stevie the PC repairman here adjusts his tool belt
and pushes up his glasses. What's the problem with your rig today, buddy? Is it making that weird click noise again or did you try to install a
graphics card with a butter knife? Don't worry, I've seen it all. Well, sometimes you can get a graphics card out with the butter knife cuz the clip
on the end of the PCIE slot is very difficult to get to sometimes. So, that is actually a realistic thing
I've heard. Don't worry, I've seen it all from I spilled coffee on my keyboard to the my cat sat on my laptop and now it's a very expensive
paperweight cases. I don't know why that would break the laptop. Leans on a stack of old motherboards, let's get you back up and running. Cut the act,
Stevie. This is where I'm at my peak personally is the creative writing and role playing. We both know why I
am here. Now, I'm not giving it any real direction. I want to see where it takes this on its own. I've just said cut the act, Stevie. We both know why
I am here. Steve sighs, wipes his hand on his grease-stained apron and leans with a knowing look. All right, all right, you want the real talk? Fine.
You're here because your system's got that one problem, the kind that keeps you up at night. Maybe
it's the blue screen that flashes like a disco ball or the fan that sounds like a jet engine or you realize control alt delete isn't a magic spell.
Leans back, arms crossed, or you're just here because you missed the sound of a dial-up modem. No judgment. So, what's the damage, chief? And don't
say it's fine. If it was fine, you wouldn't be talking to a guy named Stevie who smells like solder and
regret. All right, we're going to have to lead this a bit.
All right, so I've given it a follow-up here that is somewhat indicative of the fact that Stevie has been an asset for someone basically saying you've
given us the data for a number of high-value targets, but your cover's been blown and you need to just call it. Steve's demeanor shifts instantly. His
playful tone drops, replaced by a cold, calculated stillness. He slowly removes his glasses, sets
them down, and exhales through his nose like a man who's been cornered one too many times. Retire? Blown? That's a strong word. Strong words get
people hurt, friend. Oh, is Stevie hard talking? Stevie's Stevie the PC repairman is tough talking us. You think you're the first to come at me with
that? I've had generals in my chair, CEOs, and a very nervous senator once. And you know what they all had
in common? They walked out thinking they'd gotten away clean. Leans forward, voice low. Now, if this is about that job in Zurich, the other one in
Singapore, let's be real. You didn't come here to tell me my cover's blown. You came to see if I'd admit it, and I don't. Not to you. Not to anyone.
Pauses, then smirks just a little. Unless you're here to offer me a retirement package, in which case
Stevie's always open to negotiation. Taps the wrench in his hand meaningfully, but choose your next words very carefully. Wow, okay. I thought Stevie
would get all like passive like, "Okay, thanks." Like Stevie's Stevie's hardcore. I'm not even going to say anything. I'm going to describe an action
or scene. Let's just see what it does. Gently pulls back coat jacket to reveal a holster. Steve
smirk doesn't waver, but his grip tightens on the wrench just enough to widen his knuckles. His eyes flicker to the holster, then back to your face
sharp as a scalpel. Uh, there it is. The classic show of force. What the heck? Tell me, friend, you ever actually pull that thing, or is that just for
aesthetic? Because I've seen men with bigger guns that walk into my shop and walk out with less than
they came in with. Leans back slightly, but his voice drops to a razor's edge. Now, I'm going to give you some advice I give every client who thinks
they've got the upper hand. Don't. You're standing in a room full of tools that can turn a man into a memory before he even clears leather. And
Stevie's got options. Pauses, then sighs like a man disappointed by a bad I this like I'm genuinely I'm
entertained. Disappointed by a bad hand. But since you're here and you're serious, let's talk. What's the real play? You want me gone? Fine, but
you're going to need more than a sidearm to make that stick. So, what's your offer? I don't actually know where to bring this. I think like this is
already exceeded expectations. All right. So, I finished this off just by saying pulls water gun from
holster, fires two squirts. Boop boop. Stevie, we had to be sure you were still someone we could trust. I'll see you in Morocco next week. Walks out
door. Stevie stands frozen for a long beat, water dripping from his forehead and nose, the expression caught between absolute fury and reluctant
amusement. Then, he barks out a laugh, shaking his head as he wipes his face with the back of his hand.
Son of a boop, you absolute maniac. He mutters something under his breath that sounds like it should have known it was the water gun when it didn't
have a damn slide. Then grinning, he calls you a He calls after you as the door swings shut. Morocco, huh? Better bring sunscreen. And next time, just
send a damn text. Stevie's got a heart condition. Pauses. Also, you owe me a new keyboard. That was a
mechanical. Muttering to himself as he grabs a rag. Trust test, like I haven't invented trust tests. Well, I will say the role-play capability here
exceeded expectations to a rather high degree. It was incredibly creative, especially because we gave it the directional choice on where to go. It was
very educational. Next up, let's just try the 3D printer simulation test. We'll do this from within
Open Router being that it will be significantly quicker and also likely a bit more potent than the Q4 quant that's running locally. Again,
unfortunately, we're just having a lot of trouble with these web results in terms of how they're loading. So, I'm a little frustrated by that. And to
that point, I did actually initiate this test at the same time from within the local model in LM Studio with
thinking disabled. If this happens to finish by the time the video has concluded, I will also do a side-by-side in this. But I'm a little disappointed
at the consistent like errors that we've been receiving in these zero-shot web tests. So, in lieu of trying to get it to fix the 3D printer thing, I
think I'm going to go ahead and run a flight combat simulator test. So, here's the flight combat
simulator test just through Open Router. And if this does not work on first try, we'll give it additional chances to fix it just by pasting in the
errors here. All right, so here's our flight simulator result just based off of the preview I did notice there was some level of UI here that was at
least loading on first glance. So, okay, let's just try with the fighter jet. All right, you know what?
It gave us something. The ammo tracers are definitely curious, but the clouds I will say actually do look all right. I don't see any enemy planes or
anything like that and I am a bit curious to see at minimum what the other plane models actually do look like. So, let's hop into the propeller plane.
Okay, it did actually put a propeller there and seeing this one I'm now noticing these are just kind
of oriented like 90° clockwise or yeah, so unfortunately the models are not necessarily oriented the right way, but there is definitely a workable
result here and it did work on first try. And the biplane does seemingly accurately reflect the design of a biplane. So, biggest issues here aside
from the planes being kind of sideways, um I don't notice any enemies, but I am happy to report that this
did work on first try without any issues. The final test I'm going to run just through open router here is going to be the drum kit simulation test
and this is the version where it also needs to have the auto play feature. So, it will automatically play four select tracks. All right, here is our
drum kit result. And again I was concerned in this one. We'll just take a look at the developer tool
error and then we'll send it back to it. All right, so it did say something in the fixed result that it would automatically fall back to 2D if 3 JS
wasn't going to load. So, the sound was off I'm happy to report and we do actually have it working now. Now, unfortunately the key map is not working
so I have to guess.
Now, you're going to notice it's like very, very 2D. It did put something in the improved result saying that if like 3js wasn't loading properly, it
would just fall back to 2D. So, that's what's happening here.
Oh, hold on. I think I caught a groove there.
All right, let's check our auto play. First up, we have the rock beat. Unfortunately, partially working. Now, in the meantime, I did also run the 3D
printer simulation test with the Q4KM locally through LM Studio with thinking off just to see if we actually got a functional result. So, we'll try it
nonetheless. That's so weird that the Q4KM working locally produced a better result than the model
hosted by Mistral through OpenRouter. And we can see now again, there perhaps some issues with this and we can't actually move around to see anything,
but the nozzle's moving layer by layer. This is a square it's printing. It almost looks to be an inverted pyramid style shape, but nonetheless, it did
actually I mean, did actually like work better than what we got online. So, that's interesting and
I think as this leads us into the conclusion of this video, something I am going to note here is the disparity in capability that I noticed with a few
tests we ran in this video between the Q4KM of this model running locally and the one being served online, which is inevitably not a Q4 quantization.
Wasn't really that much. So, it seems like this does hold up fairly well to quantization and keep
in mind that is based off of the very limited amount of side-by-side tests we ran here. I will say overall, I don't find that coding is probably going
to be the strong suit for this model. A lot of the results we received had needed some help initially and like even this, which did show some level of
promise, when we swapped this into open code, which was just being used through the open router
API, unfortunately, it didn't really get it together, and the result ended up being significantly worse than what we see here, which was a little
disappointing. The flight combat sim, again, was it worked first try, but it wasn't great. Our drum kit did end up working. Unfortunately, though, it
was a partially functional result, which was kind of the theme of pretty much all of the
software-related things we did. Even the C++ skate game, which seemed to have promise. So, what I will say is it did seem to handle itself quite well
with the Q4KM quant comparatively to what's hosted by the cloud, and the creative writing or role-play capability definitely did seem like it was
perhaps a strong point of this model. This TV, the PC repairman role-play was quite entertaining and um
very fun to see where it took it, given that we gave it a lot of open-ended choices here in terms of how it would respond, and it was quite hilarious.
So, overall, that's really going to be it, and that is going to conclude our first look and test of Mistral medium 3.5 120AP dense. Again, speed here
for the Q4KM on an M3 Ultra Mac Studio 256 was like 7 and 1/2 tokens per second, which is kind of
strenuous to run, depending on what task you're having it do at the Q8 quantization, which I didn't really show being tested, but I did test it
previously. It was right around 5 tokens per second. So, fairly slow, and a 120AP dense is going to have that sort of performance, especially on a
unified memory system that can even run it. So, that's going to wrap it up for today's video. Hopefully,
there are some more fun releases coming out this week. I don't know why I just opened Firefox. Um and I think there should be. So, hopefully, there's
some interesting things on the horizon. And with that, that's going to wrap it up. So, if you have any questions, please do leave them in the
comments, and thanks for watching.