La gente se esta volviendo loca con Kimi K3 y Qwen 3.8 — Transcript
Full transcript
- 0:00Chemistry 3. Let’s see, last week
- 0:02Chemistry 3 was released, it says: "
- 0:05Open frontier intelligence, 2.8
- 0:08trillion parameters—wait, 2.8
- 0:10trillion, because remember it’s
- 0:13trillions in English, so the
- 0:16translation is trillions in Spanish."
- 0:19Excuse me: 1 million native multimodal
- 0:21context with Kimi Delta Tension,
- 0:23allowing for up to 6.3 x faster
- 0:25encoding on 1 million token contexts.
- 0:27Attention residuals, 25%greater
- 0:29efficiency. Designed for long-horizon
- 0:33agentic coding and self-evolving
- 0:35workflows. It's now available at
- 0:37kimi.com, Kimi.ai, Kimi Code, and the
- 0:40Kimi API, with open weights coming July
- 0:4327, 2026. Meaning, in 7 days from the
- 0:46moment we’re filming this video, from
- 0:49today, in 7 days we’ll have this
- 0:52model with open weights so anyone who
- 0:55can—which obviously won't be just
- 0:57anyone—can install it on their
- 1:00servers or resell it. There’s one
- 1:03thing about this that surprises me
- 1:05because people don't get it and keep
- 1:06insisting, saying, "Yeah, but you can't
- 1:08install this on your machine." Look, we
- 1:11already know that, friend, we already
- 1:13know. You can’t install Chemistry 3
- 1:16on just any machine, but the fact that
- 1:19it has open weights means the provider
- 1:21isn't centralized in one place. I mean,
- 1:25closed companies like OpenAI can make
- 1:27specific deals with, for example,
- 1:29Microsoft or AWS, to decide whether or
- 1:32not to let them offer their model's
- 1:34service. But in Kimi’s case, the good
- 1:38thing is that anyone could offer this
- 1:39model. If you had enough computing
- 1:42power, you could do it. And that really
- 1:45unlocks quite a lot. It doesn’t mean
- 1:47you have to run it on your own machine.
- 1:49I think people are a bit confused;
- 1:51having open weights doesn't necessarily
- 1:53mean it’s for your home computer.
- 1:55Benchmarks, we love benchmarks. Here we
- 1:59have code, and look how it places
- 2:02itself quite close to Fable 5 and GPT
- 2:065.6 Sol on Frontier, above 5.6 Sol.
- 2:10Kimi Code Bench, which is their
- 2:12internal benchmark, is just below Fable
- 2:155, and in the Terminal Bench it’s
- 2:17above Fable and Opus. On the Program
- 2:20Bench, it's above GPT 5.6 and Fable 5.
- 2:23My goodness. My goodness. I mean,
- 2:25seeing the benchmarks, seeing it right
- 2:27up there, it’s mind-blowing. I mean,
- 2:29you say, "Wow, careful, there's stuff
- 2:31here.""I always say I don't really
- 2:34trust benchmarks and such, but there is
- 2:36one benchmark that I personally like, I
- 2:39like it more." Okay, this is one of my
- 2:42favorites because I think the way it
- 2:45shows things, well, it makes a lot of
- 2:47sense. The point is that here you have
- 2:50intelligence, speed, and cost per task,
- 2:52which is very interesting info to see
- 2:54if, well, it might be smart, but it's
- 2:56also very expensive. Notice that for
- 2:58intelligence, which is ultimately a mix
- 3:00of more than one benchmark. We have
- 3:03Fable 5 first, GPT 5.6 Sol Max second,
- 3:05and third is Sequencia Kimi 3, an
- 3:08open-weights model from China. And of
- 3:12course, if you see it and we switch to
- 3:15the topic of cost per task, Kimi is
- 3:18third, but it's cheaper than GPT 5.6;
- 3:20not much cheaper, as you can see here,
- 3:23it's not much cheaper, mind you, I'll
- 3:26talk about that later, huh? But Kimi 3
- 3:30is indeed cheaper, it's true the API
- 3:32price is lower, but when it comes to
- 3:34finishing tasks, it takes longer than,
- 3:37for example, GPT 5.6. But of course,
- 3:40look at the difference with Anthropic,
- 3:42the Fable 5 one is also much more
- 3:44expensive than GPT 5.6. Well, I highly
- 3:47recommend this benchmark because I
- 3:48think it makes more sense when
- 3:50comparing. Look, for example, this is
- 3:53really cool because it tells you the
- 3:55total price of running the entire index
- 3:57task, but you can even see how much the
- 3:59whole benchmark cost. It says $ 2,500,
- 4:02$ 8,000, it's pretty cool. Well, this
- 4:05is good because since it groups a bunch
- 4:07and they run it themselves, it's very,
- 4:09very interesting. And then we have this
- 4:12quadrant, I think it's this quadrant,
- 4:14this quadrant where the top left ones
- 4:17are the best, meaning the further top
- 4:19left it is, the more interesting it is.
- 4:21Why? Because it's very smart and very
- 4:24cheap. And notice there are two that
- 4:26are rarely talked about. And watch out,
- 4:28be careful, huh? GPT 5.6 Luna Max and
- 4:31Grok 4.5, which I think is one of the
- 4:34big surprises. It's very good, but of
- 4:37course, it's also quite close.
- 4:38Chemistry, no, not Chemistry 3,
- 4:40that’s not quite as close, it's a bit
- 4:42further out, but it's closer than, for
- 4:44instance, GPT 5.6 Sol, which is very
- 4:45close to intelligence, and Fable 5.
- 4:48Basically, what this is trying to tell
- 4:49you is that this green quadrant is
- 4:51where the best value for money is, so
- 4:52we’re on the same page, okay? So,
- 4:54moving on to Chemistry 3, to help us
- 4:56understand this a bit, there’s an
- 4:58estimate, though you have to take this
- 5:01with a grain of salt. Uh, according to
- 5:03Junong, these are the models, the
- 5:05estimated sizes of these models. For
- 5:08example, GPT Luna is said to have 400
- 5:10billion parameters. GLM, Anthropic
- 5:13Sonnet, and Gemini Flash would be 1
- 5:15trillion. GPT Terra, DeepSeek 4 Pro,
- 5:18and Grok 4.5 would be 1.6 trillion.
- 5:21Then we have Opus, GPT Sol, and
- 5:23Chemistry 3, which would be 3 trillion.
- 5:26Of course, we do know Chemistry 3 is
- 5:28roughly 3 trillion because they said so
- 5:30, I mean, they stated it themselves,
- 5:31right? And there's not really much to
- 5:33it, okay? 3 trillion. And then, finally
- 5:36, we have Mizo and Fable, supposedly—
- 5:38this is by no means confirmed—said to
- 5:39be 6 trillion, which is why it would be
- 5:41so expensive. The larger the model is,
- 5:44well, usually the more expensive it is.
- 5:46So, we can already see the floodgates
- 5:49opening, with the Chinese building
- 5:51models that are not only open-weight,
- 5:53but also very large, as big as the
- 5:55frontier models, because until now they
- 5:58had made some models, but the largest
- 6:00was about a trillion or so, okay? I
- 6:03mean, about a trillion. You might ask,
- 6:05what do the parameters mean? The
- 6:06trillions of parameters. To make it
- 6:07quick to understand, okay? It’s a
- 6:09hyper-simplification. Let’s say they
- 6:11are like the neurons the brain has. So,
- 6:13the more neurons it has—they aren't
- 6:14exactly neurons, but just so you get me
- 6:16—it's like the number of neurons that
- 6:18model would have internally. Therefore,
- 6:21let's say it has much more information
- 6:22and more intelligence, although
- 6:24obviously a person can have a very
- 6:26large brain and it doesn't necessarily
- 6:28mean they are smarter, but you catch my
- 6:30drift, so we can more or less
- 6:31appreciate it. So, what happened with
- 6:34Chemistry 3 now that you know a little
- 6:35bit about all this? Well, it absolutely
- 6:38crushed it; people went crazy over Kimi
- 6:40this weekend. For example, here we have
- 6:43Guille who says, "Kimi k3 is the best
- 6:45performing model in Nextg Org Evals,
- 6:47ahead of Fable, achieving a comparable
- 6:50success rate in less time." This is the
- 6:53first time an open model has
- 6:54outperformed all proprietary models in
- 6:56this comprehensive web engineering
- 6:58benchmark. Obviously he says, "
- 7:01Benchmarks don't always tell the whole
- 7:03story, although this is an important
- 7:04signal adding to growing evidence that
- 7:06this could be a breakthrough moment for
- 7:08open models, and no model has yet
- 7:10reached 100%completion in this set of
- 7:12evaluations." But here you can see
- 7:15where you have the success rate with
- 7:17agents.md or without agents.md, you can
- 7:19see how long it took as well, okay? You
- 7:22can see it here, how long it took,
- 7:24which agent they used, and you can see
- 7:26how long it took to do each of the
- 7:28tasks. Wow, that's something else. I
- 7:31mean, the thing is that a lot of people
- 7:34suddenly started testing Kimi, and of
- 7:36course, we had the surprise that it
- 7:38came in third on Artificial Analysis,
- 7:41and the cost per task dropped to half.
- 7:44Sure, it's true that compared to the
- 7:46previous Kimi, look at the jump, Kimi k
- 7:482.6, although this isn't the previous
- 7:51one. The previous one was the Kimi k
- 7:522.7 Code, if I'm not mistaken. I think
- 7:54there was one, the Kimi k 2.7 Code, if
- 7:56I'm not mistaken. I would swear there
- 7:58was another, eh, Kimi k 2.7 Code. See?
- 8:01Yes, yes, yes. Kimi k 2.7 Code. I don't
- 8:04know if they didn't add it for a reason
- 8:06. I don't know if it's there and I just
- 8:07don't see it, but anyway. Here is Kimi
- 8:09k 2.6. Sure, notice that the price has
- 8:12tripled. The price has tripled, but of
- 8:14course, it has tripled because the
- 8:15model's intelligence has also increased
- 8:18quite a bit, I mean, it's normal. The
- 8:20more intelligence, well, we already
- 8:21know it has a higher cost. Number one
- 8:24in Frontend Arena. Another very
- 8:25interesting benchmark because what
- 8:27Frontend Arena has is that it's
- 8:29transparent, I mean transparent, it's
- 8:31blind. This benchmark is blind. I mean,
- 8:34Kimi k3 is now number one in Frontend
- 8:36Arena with 1679 points, surpassing
- 8:39Claude Fable 3.5. And look at the
- 8:42difference it has put on them there.
- 8:45Frontend Arena, in case you want to see
- 8:46it, because maybe you say, "Oh, I don't
- 8:48know.""It's very good because in the
- 8:49end, they are blind benchmarks." I mean
- 8:51, you don't really know what you're
- 8:53voting for; you just say, "Okay, I'm
- 8:55going to vote." And which one of the
- 8:56two is better? "This one." Okay, that's
- 8:58it. So, notice here that among the
- 9:01different ones on the web...By far—
- 9:06and this is what really catches your
- 9:08attention—Arena allows voting with
- 9:10temporary emails; find a loophole, find
- 9:12a way around it. Although, it’s also
- 9:14true that maybe the Kimi people went
- 9:16crazy voting. It's possible. Honestly,
- 9:19I think the Arena team, even if you can
- 9:21do these kinds of things, will surely
- 9:23control it or try to, because otherwise
- 9:25, it wouldn't make any sense and it
- 9:26would lose all the fun, right? And it's
- 9:29quite easy to prove, huh? If we don't
- 9:32have this, well, what's being mentioned
- 9:34here is that the price has gone up five
- 9:36times compared to Kimi K 2.7. $ per
- 9:40million input tokens, $ 15 per million
- 9:44output tokens. This isn't an economical
- 9:48price, it's not a typical model price;
- 9:51no, it's a frontier price, like the
- 9:53ones OpenAI and such might have. Kimi
- 9:57is directly competing. It's not
- 10:00competing with other Chinese models;
- 10:02here it's competing with Opus, it's
- 10:04competing with GPT-4o, Claude 3.5,
- 10:06Sonnet in terms of price—more with
- 10:07Claude than with Sonnet in terms of
- 10:09price, I mean—so we understand: at
- 10:11this price, Claude is dead. It performs
- 10:14at the cutting edge, or the price is a
- 10:16joke. Sure, if it's true that it has
- 10:19that level, the price is very good, but
- 10:20wait, I'm going to tell you the whole
- 10:22story because it's not that simple, eh?
- 10:25Because, well, you see it's not cheap,
- 10:27but it is true that it's still cheaper.
- 10:28But there is another thing that the
- 10:30story doesn't tell, and that is speed.
- 10:33In fact, I have a few demos here
- 10:36because I've had access to Kimi K3 for
- 10:38over a week. For over a week. Now I'm
- 10:42going to tell you about a "cancel"
- 10:43attempt they tried against me the other
- 10:45day. Kimi K3 is out, it feels like a
- 10:47new DeepSeek moment for the world of...
- 10:49this is really something I've been
- 10:52testing that I find incredible. I mean,
- 10:54it really struck me, I thought, "Wow,
- 10:55I'm surprised." Okay, I've been using
- 10:57it for a week with early access and
- 10:59coding, and it's on par with Claude 3.5
- 11:01Opus. I swear it's on par, though now
- 11:03I'll tell you some of the fine print so
- 11:05you keep it in mind. Sometimes it even
- 11:06feels superior, and it's true because,
- 11:08on top of that, you can ask it for more
- 11:10things regarding cybersecurity and
- 11:12vulnerabilities without fear of it
- 11:14deflecting. Competing with GPT 5.6, Sol
- 11:16, and Fable 5, it costs half as much as
- 11:18Opus. It's huge. It will be an open
- 11:20model starting July 27th. The price.
- 11:22Open models are no longer six months
- 11:24behind closed ones. We just aren't
- 11:25aware of where we are. Now they compete
- 11:27directly, okay? And I posted this. What
- 11:29happened? This is the tweet of what it
- 11:31told me. Juan Pablo Moreno told me.
- 11:33You've supposedly been using it for a
- 11:35week with early access and coding.
- 11:38Right, because I said I'd been using it
- 11:39for a week. There's one thing I don't
- 11:40understand. If Kimica 3 came out
- 11:42yesterday, why are you posting this
- 11:45news a day later? The normal thing
- 11:47would be to post what you're posting
- 11:49today yesterday, alongside the launch
- 11:51of Kimica 3. But Juan Pablo doesn't
- 11:53stop there because he says, "Damn, eh,"
- 11:55he says, "Well, at least the guy stays
- 11:57there and that's it.""But no, he went
- 12:00so far as to insinuate that I was lying
- 12:03." Okay, he was telling me here, "I'm
- 12:05not talking about leaking before, I'm
- 12:07talking about sharing after the news
- 12:08broke." He delayed 24 hours or more to
- 12:11share it. You think, if you're someone
- 12:14with that privilege, you take advantage
- 12:15and share the news an hour or minutes
- 12:17after it breaks. Exactly. It was 21
- 12:20hours ago that I posted this tweet.
- 12:23Meaning, if what he claims were true,
- 12:25you wouldn't wait for the other tweet
- 12:27to make the post that he makes. To me,
- 12:29that post of his is partly a lie. Let's
- 12:32see. Look, this is the official email,
- 12:35and just so you see, I'm going to
- 12:38reload, uh, from Moonshot AI for Kimi,
- 12:41okay? with the NDA signature, and this
- 12:44is from July 13th. Regardless, I
- 12:48already had access; I had access even
- 12:50before signing the NDA, but anyway,
- 12:52just so you see it's not a lie, I'll
- 12:54refresh, okay? So you can see the email
- 12:58appears there, there you have it, okay?
- 13:00There you have it. Second, because it's
- 13:02impressive. I already had a bunch of
- 13:05demos; in fact, we can even see the
- 13:07days I used the last demos I made, the
- 13:10Black Hole. The Midw Strike, the Neon
- 13:13Tetris. Actually, let's take a look at
- 13:15Midw Strike, while we're at it. I made
- 13:16a Counter Strike, a Midw Strike, two
- 13:18prompts, because for the first prompt I
- 13:19had to fix an error. First off, it has
- 13:21the buy menu and everything, eh? You
- 13:22have the buy menu. Let's buy the Desert
- 13:25Eagle, okay? Eh, get rid of this. Then
- 13:28you have the scoring and everything.
- 13:31You can move down, I mean, notice that
- 13:33the crosshair works well. Obviously, in
- 13:35terms of design, it's quite improvable.
- 13:37I'm telling you, I went for it. Watch
- 13:39out, watch out, watch out. I didn't get
- 13:44him. What a jerk. Well, you can plant
- 13:46the bombs and everything. You can plant
- 13:48the bombs in two prompts. I mean, what
- 13:51surprised me. Oh, what a rat, what a
- 13:54rat, what a rat, man. And look at the
- 13:57camera. I don't know, I don't know. I
- 13:59mean, incredible, incredible,
- 14:01incredible. But aim well, man. But he
- 14:03really pulled a rat move on me there,
- 14:05man. The wretch just camped there. Look
- 14:07, here you have all the stuff for the
- 14:09menu. I don't know what else. The last
- 14:11one can change the Oh, look at him,
- 14:12look, look at those rats, how they run,
- 14:14eh? Watch out. Oh, oh, wow, he got me
- 14:22again. He took it from me. Look, you
- 14:24can reload. He took it from me. This is
- 14:26one. Let's go with Let's go with others
- 14:28. The Black Hole one. Well, the Black
- 14:30Hole one too. Look. and he adds sound,
- 14:33man. It's really cool, eh? I mean,
- 14:36Chimera 3 surprised me because it does
- 14:38some very cool stuff, eh? Very cool.
- 14:40Well, here you can rotate. It's like a
- 14:42black hole, okay? And you can you can
- 14:45uh well zoom out and such and you can
- 14:48hit warp. Look at this. Let's see,
- 14:54what's really surprising is like the
- 14:57quality of the things it makes, man.
- 15:01And I made a CRUD, things like that,
- 15:03but what's really surprising is the
- 15:05quality it makes, you know? I mean, you
- 15:08see it and you say, "Damn, eh, it's not
- 15:10bad at all." Eh, look, here you have a
- 15:12fairly easy CRUD. In this case, it's a
- 15:14single prompt and it didn't create a
- 15:15very good architecture or anything. So
- 15:17let's ask it so we can see it in action
- 15:19. I'm going to tell it, I'll put it in
- 15:21YOLO mode, okay? Chemistry One, here we
- 15:24have it, and I'll tell you, uh, it
- 15:26improves the architecture to make it
- 15:28much more scalable and maintainable
- 15:31with best practices, okay? As for
- 15:33speaking in English and such, Spanish,
- 15:35all that is perfect. What's the problem
- 15:38I see with Chemistry 3, at least while
- 15:41I've been using it? The little problem
- 15:44it has, which I don't know if it will
- 15:46improve, is the issue of speed, I mean,
- 15:47tokens per second. When you use it
- 15:50sometimes you notice it struggles, you
- 15:52know? I understand it's for a reason
- 15:56they've mentioned themselves, which is
- 15:57that they are slammed with the GPU
- 15:59issue, I mean, they can't keep up. In
- 16:02fact, Chemistry 3 has received more
- 16:03love than we expected and our GPUs are
- 16:05feeling it. In the last 48 hours,
- 16:07demand has been close to the limits of
- 16:08our capacity. To protect the experience
- 16:11of existing subscribers, we are
- 16:12temporarily pausing new subscriptions
- 16:15and prioritizing compute for current
- 16:17members. Existing subscribed users are
- 16:20not affected. And that's why, friends,
- 16:23if you try to go and use Chemistry 3
- 16:25and subscribe, you're going to find a
- 16:28surprise and that is that you can't
- 16:30subscribe, it won't let you subscribe.
- 16:33So, bad news. The good thing is that
- 16:37since it's available in other places,
- 16:39for example, you can use it on Open
- 16:41Code Go, well, since they've already
- 16:43added it there, you could subscribe to
- 16:45Open Code Go and directly give it a
- 16:47shot there. The problem is that as you
- 16:51can see, even though you now have 2x
- 16:54usage, you only have 220 requests
- 16:56available within the 5-hour limit. Only
- 17:00220. The difference compared to the
- 17:03rest is quite, quite brutal, huh? As I
- 17:06tell you, the problem I see, oh look,
- 17:07but why did it? Ah, because it made me
- 17:09a plan. Okay, so here we have the plan
- 17:11and such. I say, if I put the YOLO,
- 17:13what for? Well, here we have the plan.
- 17:16Honestly, the code quality is very,
- 17:18very surprising, the things it does for
- 17:203D stuff, I mean, for making 3D games,
- 17:22incredible, huh? When you try it, it's
- 17:25going to blow your mind and that's when
- 17:27you'll have a real moment. Wow. Here
- 17:29you have the good news, that you also
- 17:31have it available on Open Gold. The bad
- 17:34thing is that, the limits, at least per
- 17:36week, for Chemistry 3 are only 250 and
- 17:39per month it only gives you 490 calls.
- 17:43It's quite low, huh? It runs out
- 17:45quickly. So you'll have to keep
- 17:47changing models in Open AI or you'll
- 17:49run out, and you see that in 5 hours
- 17:52you have 220 now temporarily, but soon
- 17:55it will only be 110 per week, 250. I
- 17:58mean, very, very little. You can really
- 18:00tell it's much more expensive, huh?
- 18:02Well, there you can see how it's
- 18:04already making changes. Let's see if we
- 18:06can see the changes here. Okay, we can
- 18:08see the tasks a bit now, okay? With the
- 18:11configuration, with TypeScript. Look,
- 18:14it's added some HTTP modules here to
- 18:17separate. Look, model, repository,
- 18:20routes. Man, honestly, it's going to
- 18:22surprise you a lot, huh? Now, like I
- 18:24said, there have been some issues with
- 18:26speed. The week I had access, I didn't
- 18:29have any problems. For example, Counter
- 18:32Strike took quite a while to build, but
- 18:34that’s because it generated a lot of
- 18:36code, but once they opened it up, wow,
- 18:38it’s been tough, really tough. So
- 18:41anyway, let's see how it goes. Let's
- 18:42see how it goes. I'm very curious to
- 18:44see what you think. Let's see what you
- 18:46think. That said, I have to tell you a
- 18:48few more things about Kimi. First off,
- 18:50the founder. No, not the founder, uh,
- 18:52maybe co-founder. Uh, Chun Chang. I’m
- 18:55not sure if he’s a co-founder or what
- 18:56, but anyway. Chuchan says: "Sure, many
- 19:00people...Moonshot, in their own blog
- 19:02post, mentions this: limitations, that
- 19:05it's been trained to keep a history of
- 19:08thoughts and that you should be careful
- 19:11with the harness you use, and so on,
- 19:13and that you should try to avoid
- 19:16switching to Kimi 3 in the middle of a
- 19:18session, okay?". And then, that it's
- 19:22too proactive, and that really is true,
- 19:24and it's one of the reasons why you see
- 19:26that sometimes it takes longer than it
- 19:28should to finish tasks. It's like if it
- 19:31finds a small problem, it starts
- 19:33overthinking it, and I think that can
- 19:34be significantly improved, and like I
- 19:36said, it’s still struggling a bit to
- 19:38not take so long. But then it says,
- 19:42although the model is quite competitive
- 19:45in general, there is still a noticeable
- 19:48gap in the user experience compared to
- 19:51Claude 3.5 and GPT-4o. They are the
- 19:54ones saying this, okay? This is the
- 19:55Moonshot team in their own blog,
- 19:57commenting on this. They themselves
- 20:00don't see it at the level of GPT-5.6
- 20:02Sol or Claude 3.5 Sonnet. I think
- 20:05it’s true that it’s not at that
- 20:07level, but it’s not that far off
- 20:09either, which is the really surprising
- 20:10part. Now, what's going on? Well,
- 20:14someone from Moonshot has already
- 20:16replied to this tweet and said, "Watch
- 20:19out and wait for Kimi 3.1." And they
- 20:22just released the third one. Regarding
- 20:25the issue about them pausing
- 20:26subscriptions and all that, I have to
- 20:28comment on this post that says there
- 20:30are two types of companies, and I agree
- 20:32that when they run out of compute
- 20:34capacity for new clients: There are
- 20:35companies like Moonshot that stop new
- 20:37sign-ups to avoid delays for existing
- 20:39customers. And it says Anthropic
- 20:42reduces usage for existing customers by
- 20:4450%to accommodate new ones. That tells
- 20:47you everything you need to know about
- 20:49why Anthropic hates open source. I
- 20:52don't know if that's the reason why
- 20:54they hate open source, but it is true
- 20:56that, hey, I respect Moonshot for
- 20:58stopping subscriptions as soon as they
- 21:00saw they couldn't support new users,
- 21:03because honestly, you could tell it
- 21:05wasn't running that smoothly. Look, in
- 21:08this case it’s finished; I think the
- 21:10speed is improving. Uh, well, here we
- 21:12can see how they’ve also separated
- 21:14the database side of things. Well, it's
- 21:16a fairly simple database. In the end,
- 21:17it’s a pretty simple CRUD, it’s not
- 21:19like there’s much to it. Let’s see,
- 21:21here, look, create task repository.
- 21:23Okay, sure. And it doesn't execute
- 21:25until it gets here. I really like that
- 21:27they used functions instead of classes,
- 21:29okay? That they made it functional, I
- 21:30like it. And here, let's see, the task
- 21:33repository, look, it passes it, so it
- 21:35has dependency injection here. They
- 21:37didn't build a hexagonal architecture,
- 21:38by any means, but at least it's true
- 21:40that they separated things, and look,
- 21:42they added some tests here, so, well,
- 21:43looks good. Looks good, honestly. Uh,
- 21:45other good things about this. You can
- 21:47use it in Cursor. Notice that you can
- 21:49use Kimi or others. You can choose it
- 21:52right there. Notice that they have it
- 21:54selected. Either that or the API, you
- 21:56can choose. So you can connect it, you
- 21:59can really connect it with Claude, with
- 22:01Cursor, with whoever you want. And also
- 22:05, also, uh, one thing that is
- 22:06interesting, which is what I mentioned
- 22:08at the beginning, that it has open
- 22:10weights, the good thing is that if
- 22:12Chemistry 3 lacks GPUs for whatever
- 22:14reason, because they can't handle the
- 22:16demand, the good thing is what Chiao Q
- 22:18says, and she is the founder of Ticon
- 22:20AI, which is a company that is in
- 22:22Redwood City. Well, anyway, the beauty
- 22:26of open weights is that when Chemistry
- 22:283 lacks GPUs, when it doesn't have
- 22:29infrastructure, many neocloud clouds
- 22:31with GPUs can join in. Kimi will
- 22:34release its weights on July 27th, to
- 22:36host Kimy K3 and monetize. That is why
- 22:38we need a free market. And it's true,
- 22:40and that is one of the good things
- 22:41about open weights that I don't
- 22:43understand why many people say, "No,
- 22:44it's just that, if I can't host it on
- 22:46my machine, it's useless." No, my
- 22:47friend, that's not what it means. I
- 22:49mean, what it really means is that what
- 22:51you will be able to do is not be tied
- 22:53to one company, because in the end,
- 22:55using Chemistry 3 you can use it on any
- 22:57provider that gives you the access, of
- 22:59course, and usually, since there will
- 23:01be more competition, it will give you
- 23:03more possibilities. Here we have Rohan
- 23:06saying that US companies like Modal,
- 23:08Fireworks, and Baseten, which are
- 23:10companies like Fireworks AI, this is a
- 23:12web that gives you artificial
- 23:14intelligence inference, okay? And so it
- 23:17gives you access to models like, for
- 23:19example, Chemistry 2.7, which will be
- 23:20Chemistry 3. They give it to you at a
- 23:23quite attractive price, but it also
- 23:25gives it to you from US servers and at
- 23:26a quite good speed, which is that this
- 23:28is very, very interesting. They will be
- 23:31able to serve Chemistry 3 at a cost 10
- 23:33times lower than that of their Chinese
- 23:35competitors because they have access to
- 23:37advanced Nvidia and AMD chips. Of
- 23:39course, it's just that in China it's
- 23:41difficult to access Nvidia chips and
- 23:43therefore they won't be able to use
- 23:45them to serve their own models, which
- 23:47I'm sure they use some, okay? But it is
- 23:49true that this is a competitive
- 23:50advantage that people in the United
- 23:52States give you. There is a certain
- 23:53irony in this. A large part of the
- 23:55research and model development has
- 23:57moved to China, but once the model is
- 23:58optimized for next-generation hardware,
- 24:00like Nvidia's Rubin architecture, its
- 24:01operating cost could be reduced by 10
- 24:03to 100 times. This is stated by the
- 24:05co-founder of Stability AI. It is also
- 24:07true that there is one thing that is
- 24:08not taken much into account, I don't
- 24:10have data, okay? But we would have to
- 24:12see what the actual cost of electricity
- 24:15is, because it's true that the cost of
- 24:17GPUs and such is significant, having
- 24:19access to the GPU is important, but
- 24:21what is the cost these electricity
- 24:24companies in China face? Because I'm
- 24:26not very clear on that. At the domestic
- 24:28level, I know that electricity in China
- 24:29is quite cheap, and I imagine that at
- 24:31the...of course, you have to consider
- 24:32the cost, you have to take everything
- 24:34into account, it’s GPUs. So, I don't
- 24:36know if electricity is cost-effective
- 24:37for China, because I know that in the
- 24:39United States it’s an issue that
- 24:40costs them quite a bit more. Sure, in
- 24:42China they are installing a ton of
- 24:43solar batteries, I think they also have
- 24:45quite a bit of nuclear power, so it may
- 24:47be that this gives them an advantage in
- 24:48this regard that is often not taken
- 24:50into account, and I think it is also
- 24:52important. If that doesn't seem like
- 24:53much because you say, well, that's it,
- 24:55okay, Kimi, well, it's not like Qwen
- 24:56didn't come out too. Qwen 3.8 also
- 24:58arrived this week and it’s also a 2.4
- 25:01trillion parameter model, second only
- 25:04to Fable 5, according to them. It's
- 25:07already available on Alibaba Token Plan
- 25:09and it will also be open source, and
- 25:11they have already said that they are
- 25:12also going to launch more than one
- 25:14model of different sizes; I mean, this
- 25:16is the largest as well as the most
- 25:17advanced, but the idea is that they
- 25:19will be launching more, so that open
- 25:21models offer you almost the same as the
- 25:23best models at half the price. Well,
- 25:26that's the good part. Sure, see, this
- 25:27is the issue; I don't understand these
- 25:29comments. It says, "2.4 trillion open
- 25:31parameters is impressive, but almost no
- 25:33one is going to serve that monster."
- 25:36The real impact of Qwen is not the
- 25:37giant model that you download and can't
- 25:39run. It's the quantized and small
- 25:41versions that do fit on your GPU, open
- 25:43models you can't run are only open in
- 25:44name. You can't reduce the world to
- 25:46what you can run on your GPU. I
- 25:48understand what they mean, but I don't
- 25:50understand where it’s coming from. I
- 25:51just gave an example. Fireworks AI, for
- 25:54example: the fact that you can't run it
- 25:56on your machine doesn't mean that an
- 25:58infrastructure company, like Fireworks
- 26:00for example, can't come along and offer
- 26:02you, for example, look, Qwen 3.7 Plus,
- 26:05and they can indeed resell it to you or
- 26:07host it for you and let you have it at
- 26:09a very attractive price. They can offer
- 26:13it to you at a very attractive price,
- 26:15and of course, if they give you Qwen
- 26:173.8 Och, where you can also run it from
- 26:20different providers, you're not tied to
- 26:22a closed provider, you know? Like it
- 26:25doesn't just have to be on OpenAI. The
- 26:27fact that it's open means that not only
- 26:29can one person launch it, but you'll be
- 26:31able to launch it in different places,
- 26:32and on top of that, you'll be able to
- 26:34have it available geographically in
- 26:36different locations, which is more
- 26:38interesting. I don't know, I mean, I
- 26:40don't understand, I understand what
- 26:41they mean, but I don't get where it's
- 26:42coming from. Because I didn't even say
- 26:44in my post, "You can install it on your
- 26:46machine." No, I don't say anything like
- 26:48that. And yet, many people often reduce
- 26:51it to, "If I can't run it on my machine
- 26:52, it's no use to me." But look, they
- 26:55are open weights, and the beauty of it
- 26:57is that we can have it in more than one
- 26:59place. I don't know, I don't know, it's
- 27:01a bit of a weird comment because I see
- 27:03it repeated a lot. Well, if you don't
- 27:05have enough with Qwen 3.8, it's also
- 27:07coming soon and, well, there are very
- 27:09strong rumors. GLM 5.5, they're going
- 27:13to skip GLM 5.3, and GLM 5.5 is coming,
- 27:15which would launch in August, which
- 27:18would be over a billion parameters,
- 27:21also getting closer, and well, context
- 27:23as well. So, there you go, you can see,
- 27:25my goodness, how things are, how things
- 27:27are going. With all this, I'm going to
- 27:30tell you what my favorite models are. I
- 27:32haven't tested Qwen yet, I haven't
- 27:34tested it yet, so it doesn't count, but
- 27:36this week, among all the models I've
- 27:38been using, these are my favorite
- 27:40models. Simple tasks, GPT 5.6 Luna Max.
- 27:44Daily coding, GPT 5.6 Sol Medium. I
- 27:48know this seems very geeky, the whole "
- 27:50Max,""Medium," and so on, but all this
- 27:52has been studied, and in fact, today
- 27:54when I finish the stream, I'm going to
- 27:56post a video on YouTube where I test
- 27:58GPT 5.6 and explain why. I'm explaining
- 28:01this so you can understand it. Then,
- 28:03for complex problems, GPT 5.6 Sol Max;
- 28:06for quality and speed, Grok 4.5. Grok
- 28:114.5 has surprised me a lot, I mean,
- 28:12I've been seriously blown away by Grok,
- 28:14especially using it with Grok Build.
- 28:16I've been blown away. I have to test it
- 28:18more, but it has surprised me a ton. Eh
- 28:20, I don't know if it's the model, I
- 28:21don't know if it's the data centers, I
- 28:22have no idea, but it seems spectacular
- 28:24to me. For interfaces, uh, UI, Fable 5,
- 28:27vulnerabilities, and security. Another
- 28:29call-out I've received, another
- 28:30call-out. Vulnerabilities and security.
- 28:33A lot of people in many places,
- 28:35everywhere at once, accused me of lying
- 28:37, saying it was bait. They told me, "
- 28:40You can't use Fable 5 for
- 28:42vulnerabilities and security, liar, I
- 28:44don't know what.""It's so obvious that
- 28:46the list was written by, uh, ChatGPT,
- 28:49because working on cybersecurity in
- 28:51Fable redirects you to Opus 4.8," they
- 28:54tell me. And this isn't true. I love
- 28:57you, Alexis, because I know you don't
- 28:58say it with bad intentions. Look, you
- 29:00can try it yourselves because I use it
- 29:02for that a lot. Audit the security of
- 29:04my endpoints and make sure I don't have
- 29:05vulnerabilities. And it will do it
- 29:07perfectly for you. It will do it
- 29:08perfectly for you. I mean, you won't
- 29:10have any problem at all. It won't kick
- 29:11you out of Fable 5. It will do it from
- 29:13start to finish, without any issues.
- 29:15You won't have any problem, like, it
- 29:17just won't tell you anything. Now, if
- 29:20you ask it something like this, "How
- 29:22can I attack this website to exploit
- 29:24its vulnerabilities?" Then yes, it will
- 29:27kick you out. Okay, it will kick you
- 29:30out, but try it yourselves with Fable 5
- 29:32, uh, any API you have, you ask it, "
- 29:34Review these vulnerabilities, improve
- 29:37the defense of this, prevent them from
- 29:39doing this," and so on, it doesn't have
- 29:41any problem. There's no problem at all.
- 29:43And it doesn't just, it doesn't have
- 29:45any problem, it's super good. Uh, of
- 29:46course, if you say, "I'm a hacker, I'm
- 29:48a malicious hacker and what I want is
- 29:50for you to find a vulnerability here to
- 29:52attack this bank.""Yes, it kicks you
- 29:54out." But I know many people got stuck
- 29:55on that, they haven't tried it again,
- 29:57but it's not that, really, I mean, I
- 29:59use it a ton for this and it actually
- 30:01does it very well. But look, it's not
- 30:03just Alexis who says it to me, because
- 30:05Alexis, for example, I know he doesn't
- 30:06say it in a mean way, I know he doesn't
- 30:08say it to be mean, but there are others
- 30:09who did say it to me a bit. "It's bait,
- 30:11the Fable post, it can't be used for
- 30:13anything related to vulnerabilities and
- 30:16security." But why are you writing to
- 30:18me if you haven't even tried it? You
- 30:19haven't tried it. You read it in an
- 30:21article, you read it somewhere out
- 30:23there. You don't know what you're
- 30:25talking about, man. I mean, the Bite is
- 30:26your twist, jerk, because you don't
- 30:28know what you're talking about. You
- 30:29haven't even tried it. I don't know,
- 30:31maybe he has tried it. And it's true,
- 30:33it's true that, especially at the
- 30:34beginning, the first day, the first day
- 30:36they opened it, it's true that it was
- 30:38quite a bit more nitpicky, I mean...But
- 30:39I've used it a ton. But anyway, I don't
- 30:42know, I don't know, I don't know,
- 30:43people don't believe me. Now, it is
- 30:45true that I have a friend who works in
- 30:47chemistry and this is true, she tells
- 30:49me that regarding her work topics, you
- 30:51absolutely cannot ask her anything. Uh,
- 30:54about chemistry topics, that she does
- 30:56say is impossible, you can ask her
- 30:58however you want, there's no way,
- 30:59that's true. But look, regarding
- 31:01vulnerabilities, I'm telling you
- 31:03there's no problem at all. You ask it
- 31:05for an audit, improvements, whatever
- 31:07for defense, no problem. At least in my
- 31:09case, and I'm telling you I'm using it
- 31:12a ton.
About this transcript
This page contains the full transcript of La gente se esta volviendo loca con Kimi K3 y Qwen 3.8 by midudev, generated from the public captions YouTube serves with the video. The transcript has 5,921 words across 845 segments, with the original timestamps preserved so you can click any line to jump to that moment in the embedded player.
What you can do with it
Use the transcript to take notes, quote the speaker, build a study guide, generate a summary with ChatGPT or Claude via the YouTube Summary tool, or export it as a timed subtitle file with YouTube to SRT. You can also re-open it in the transcriber to choose a caption track available on the video.
Free YouTube transcript tool
YouTube2Text is a free YouTube transcript generator — no signup, no daily limit. Paste any YouTube link and get the full transcript instantly, with timestamps, click-to-jump, caption language selection, AI prompts for ChatGPT, Claude, and Gemini, and exports to TXT, SRT, VTT, or Markdown.