YouTube2Text

La gente se esta volviendo loca con Kimi K3 y Qwen 3.8 — Transcript

by midudev · 5,921 words · 845 segments · language en · Watch on YouTube

Full transcript

  1. 0:00Chemistry 3. Let’s see, last week
  2. 0:02Chemistry 3 was released, it says: "
  3. 0:05Open frontier intelligence, 2.8
  4. 0:08trillion parameters—wait, 2.8
  5. 0:10trillion, because remember it’s
  6. 0:13trillions in English, so the
  7. 0:16translation is trillions in Spanish."
  8. 0:19Excuse me: 1 million native multimodal
  9. 0:21context with Kimi Delta Tension,
  10. 0:23allowing for up to 6.3 x faster
  11. 0:25encoding on 1 million token contexts.
  12. 0:27Attention residuals, 25%greater
  13. 0:29efficiency. Designed for long-horizon
  14. 0:33agentic coding and self-evolving
  15. 0:35workflows. It's now available at
  16. 0:37kimi.com, Kimi.ai, Kimi Code, and the
  17. 0:40Kimi API, with open weights coming July
  18. 0:4327, 2026. Meaning, in 7 days from the
  19. 0:46moment we’re filming this video, from
  20. 0:49today, in 7 days we’ll have this
  21. 0:52model with open weights so anyone who
  22. 0:55can—which obviously won't be just
  23. 0:57anyone—can install it on their
  24. 1:00servers or resell it. There’s one
  25. 1:03thing about this that surprises me
  26. 1:05because people don't get it and keep
  27. 1:06insisting, saying, "Yeah, but you can't
  28. 1:08install this on your machine." Look, we
  29. 1:11already know that, friend, we already
  30. 1:13know. You can’t install Chemistry 3
  31. 1:16on just any machine, but the fact that
  32. 1:19it has open weights means the provider
  33. 1:21isn't centralized in one place. I mean,
  34. 1:25closed companies like OpenAI can make
  35. 1:27specific deals with, for example,
  36. 1:29Microsoft or AWS, to decide whether or
  37. 1:32not to let them offer their model's
  38. 1:34service. But in Kimi’s case, the good
  39. 1:38thing is that anyone could offer this
  40. 1:39model. If you had enough computing
  41. 1:42power, you could do it. And that really
  42. 1:45unlocks quite a lot. It doesn’t mean
  43. 1:47you have to run it on your own machine.
  44. 1:49I think people are a bit confused;
  45. 1:51having open weights doesn't necessarily
  46. 1:53mean it’s for your home computer.
  47. 1:55Benchmarks, we love benchmarks. Here we
  48. 1:59have code, and look how it places
  49. 2:02itself quite close to Fable 5 and GPT
  50. 2:065.6 Sol on Frontier, above 5.6 Sol.
  51. 2:10Kimi Code Bench, which is their
  52. 2:12internal benchmark, is just below Fable
  53. 2:155, and in the Terminal Bench it’s
  54. 2:17above Fable and Opus. On the Program
  55. 2:20Bench, it's above GPT 5.6 and Fable 5.
  56. 2:23My goodness. My goodness. I mean,
  57. 2:25seeing the benchmarks, seeing it right
  58. 2:27up there, it’s mind-blowing. I mean,
  59. 2:29you say, "Wow, careful, there's stuff
  60. 2:31here.""I always say I don't really
  61. 2:34trust benchmarks and such, but there is
  62. 2:36one benchmark that I personally like, I
  63. 2:39like it more." Okay, this is one of my
  64. 2:42favorites because I think the way it
  65. 2:45shows things, well, it makes a lot of
  66. 2:47sense. The point is that here you have
  67. 2:50intelligence, speed, and cost per task,
  68. 2:52which is very interesting info to see
  69. 2:54if, well, it might be smart, but it's
  70. 2:56also very expensive. Notice that for
  71. 2:58intelligence, which is ultimately a mix
  72. 3:00of more than one benchmark. We have
  73. 3:03Fable 5 first, GPT 5.6 Sol Max second,
  74. 3:05and third is Sequencia Kimi 3, an
  75. 3:08open-weights model from China. And of
  76. 3:12course, if you see it and we switch to
  77. 3:15the topic of cost per task, Kimi is
  78. 3:18third, but it's cheaper than GPT 5.6;
  79. 3:20not much cheaper, as you can see here,
  80. 3:23it's not much cheaper, mind you, I'll
  81. 3:26talk about that later, huh? But Kimi 3
  82. 3:30is indeed cheaper, it's true the API
  83. 3:32price is lower, but when it comes to
  84. 3:34finishing tasks, it takes longer than,
  85. 3:37for example, GPT 5.6. But of course,
  86. 3:40look at the difference with Anthropic,
  87. 3:42the Fable 5 one is also much more
  88. 3:44expensive than GPT 5.6. Well, I highly
  89. 3:47recommend this benchmark because I
  90. 3:48think it makes more sense when
  91. 3:50comparing. Look, for example, this is
  92. 3:53really cool because it tells you the
  93. 3:55total price of running the entire index
  94. 3:57task, but you can even see how much the
  95. 3:59whole benchmark cost. It says $ 2,500,
  96. 4:02$ 8,000, it's pretty cool. Well, this
  97. 4:05is good because since it groups a bunch
  98. 4:07and they run it themselves, it's very,
  99. 4:09very interesting. And then we have this
  100. 4:12quadrant, I think it's this quadrant,
  101. 4:14this quadrant where the top left ones
  102. 4:17are the best, meaning the further top
  103. 4:19left it is, the more interesting it is.
  104. 4:21Why? Because it's very smart and very
  105. 4:24cheap. And notice there are two that
  106. 4:26are rarely talked about. And watch out,
  107. 4:28be careful, huh? GPT 5.6 Luna Max and
  108. 4:31Grok 4.5, which I think is one of the
  109. 4:34big surprises. It's very good, but of
  110. 4:37course, it's also quite close.
  111. 4:38Chemistry, no, not Chemistry 3,
  112. 4:40that’s not quite as close, it's a bit
  113. 4:42further out, but it's closer than, for
  114. 4:44instance, GPT 5.6 Sol, which is very
  115. 4:45close to intelligence, and Fable 5.
  116. 4:48Basically, what this is trying to tell
  117. 4:49you is that this green quadrant is
  118. 4:51where the best value for money is, so
  119. 4:52we’re on the same page, okay? So,
  120. 4:54moving on to Chemistry 3, to help us
  121. 4:56understand this a bit, there’s an
  122. 4:58estimate, though you have to take this
  123. 5:01with a grain of salt. Uh, according to
  124. 5:03Junong, these are the models, the
  125. 5:05estimated sizes of these models. For
  126. 5:08example, GPT Luna is said to have 400
  127. 5:10billion parameters. GLM, Anthropic
  128. 5:13Sonnet, and Gemini Flash would be 1
  129. 5:15trillion. GPT Terra, DeepSeek 4 Pro,
  130. 5:18and Grok 4.5 would be 1.6 trillion.
  131. 5:21Then we have Opus, GPT Sol, and
  132. 5:23Chemistry 3, which would be 3 trillion.
  133. 5:26Of course, we do know Chemistry 3 is
  134. 5:28roughly 3 trillion because they said so
  135. 5:30, I mean, they stated it themselves,
  136. 5:31right? And there's not really much to
  137. 5:33it, okay? 3 trillion. And then, finally
  138. 5:36, we have Mizo and Fable, supposedly—
  139. 5:38this is by no means confirmed—said to
  140. 5:39be 6 trillion, which is why it would be
  141. 5:41so expensive. The larger the model is,
  142. 5:44well, usually the more expensive it is.
  143. 5:46So, we can already see the floodgates
  144. 5:49opening, with the Chinese building
  145. 5:51models that are not only open-weight,
  146. 5:53but also very large, as big as the
  147. 5:55frontier models, because until now they
  148. 5:58had made some models, but the largest
  149. 6:00was about a trillion or so, okay? I
  150. 6:03mean, about a trillion. You might ask,
  151. 6:05what do the parameters mean? The
  152. 6:06trillions of parameters. To make it
  153. 6:07quick to understand, okay? It’s a
  154. 6:09hyper-simplification. Let’s say they
  155. 6:11are like the neurons the brain has. So,
  156. 6:13the more neurons it has—they aren't
  157. 6:14exactly neurons, but just so you get me
  158. 6:16—it's like the number of neurons that
  159. 6:18model would have internally. Therefore,
  160. 6:21let's say it has much more information
  161. 6:22and more intelligence, although
  162. 6:24obviously a person can have a very
  163. 6:26large brain and it doesn't necessarily
  164. 6:28mean they are smarter, but you catch my
  165. 6:30drift, so we can more or less
  166. 6:31appreciate it. So, what happened with
  167. 6:34Chemistry 3 now that you know a little
  168. 6:35bit about all this? Well, it absolutely
  169. 6:38crushed it; people went crazy over Kimi
  170. 6:40this weekend. For example, here we have
  171. 6:43Guille who says, "Kimi k3 is the best
  172. 6:45performing model in Nextg Org Evals,
  173. 6:47ahead of Fable, achieving a comparable
  174. 6:50success rate in less time." This is the
  175. 6:53first time an open model has
  176. 6:54outperformed all proprietary models in
  177. 6:56this comprehensive web engineering
  178. 6:58benchmark. Obviously he says, "
  179. 7:01Benchmarks don't always tell the whole
  180. 7:03story, although this is an important
  181. 7:04signal adding to growing evidence that
  182. 7:06this could be a breakthrough moment for
  183. 7:08open models, and no model has yet
  184. 7:10reached 100%completion in this set of
  185. 7:12evaluations." But here you can see
  186. 7:15where you have the success rate with
  187. 7:17agents.md or without agents.md, you can
  188. 7:19see how long it took as well, okay? You
  189. 7:22can see it here, how long it took,
  190. 7:24which agent they used, and you can see
  191. 7:26how long it took to do each of the
  192. 7:28tasks. Wow, that's something else. I
  193. 7:31mean, the thing is that a lot of people
  194. 7:34suddenly started testing Kimi, and of
  195. 7:36course, we had the surprise that it
  196. 7:38came in third on Artificial Analysis,
  197. 7:41and the cost per task dropped to half.
  198. 7:44Sure, it's true that compared to the
  199. 7:46previous Kimi, look at the jump, Kimi k
  200. 7:482.6, although this isn't the previous
  201. 7:51one. The previous one was the Kimi k
  202. 7:522.7 Code, if I'm not mistaken. I think
  203. 7:54there was one, the Kimi k 2.7 Code, if
  204. 7:56I'm not mistaken. I would swear there
  205. 7:58was another, eh, Kimi k 2.7 Code. See?
  206. 8:01Yes, yes, yes. Kimi k 2.7 Code. I don't
  207. 8:04know if they didn't add it for a reason
  208. 8:06. I don't know if it's there and I just
  209. 8:07don't see it, but anyway. Here is Kimi
  210. 8:09k 2.6. Sure, notice that the price has
  211. 8:12tripled. The price has tripled, but of
  212. 8:14course, it has tripled because the
  213. 8:15model's intelligence has also increased
  214. 8:18quite a bit, I mean, it's normal. The
  215. 8:20more intelligence, well, we already
  216. 8:21know it has a higher cost. Number one
  217. 8:24in Frontend Arena. Another very
  218. 8:25interesting benchmark because what
  219. 8:27Frontend Arena has is that it's
  220. 8:29transparent, I mean transparent, it's
  221. 8:31blind. This benchmark is blind. I mean,
  222. 8:34Kimi k3 is now number one in Frontend
  223. 8:36Arena with 1679 points, surpassing
  224. 8:39Claude Fable 3.5. And look at the
  225. 8:42difference it has put on them there.
  226. 8:45Frontend Arena, in case you want to see
  227. 8:46it, because maybe you say, "Oh, I don't
  228. 8:48know.""It's very good because in the
  229. 8:49end, they are blind benchmarks." I mean
  230. 8:51, you don't really know what you're
  231. 8:53voting for; you just say, "Okay, I'm
  232. 8:55going to vote." And which one of the
  233. 8:56two is better? "This one." Okay, that's
  234. 8:58it. So, notice here that among the
  235. 9:01different ones on the web...By far—
  236. 9:06and this is what really catches your
  237. 9:08attention—Arena allows voting with
  238. 9:10temporary emails; find a loophole, find
  239. 9:12a way around it. Although, it’s also
  240. 9:14true that maybe the Kimi people went
  241. 9:16crazy voting. It's possible. Honestly,
  242. 9:19I think the Arena team, even if you can
  243. 9:21do these kinds of things, will surely
  244. 9:23control it or try to, because otherwise
  245. 9:25, it wouldn't make any sense and it
  246. 9:26would lose all the fun, right? And it's
  247. 9:29quite easy to prove, huh? If we don't
  248. 9:32have this, well, what's being mentioned
  249. 9:34here is that the price has gone up five
  250. 9:36times compared to Kimi K 2.7. $ per
  251. 9:40million input tokens, $ 15 per million
  252. 9:44output tokens. This isn't an economical
  253. 9:48price, it's not a typical model price;
  254. 9:51no, it's a frontier price, like the
  255. 9:53ones OpenAI and such might have. Kimi
  256. 9:57is directly competing. It's not
  257. 10:00competing with other Chinese models;
  258. 10:02here it's competing with Opus, it's
  259. 10:04competing with GPT-4o, Claude 3.5,
  260. 10:06Sonnet in terms of price—more with
  261. 10:07Claude than with Sonnet in terms of
  262. 10:09price, I mean—so we understand: at
  263. 10:11this price, Claude is dead. It performs
  264. 10:14at the cutting edge, or the price is a
  265. 10:16joke. Sure, if it's true that it has
  266. 10:19that level, the price is very good, but
  267. 10:20wait, I'm going to tell you the whole
  268. 10:22story because it's not that simple, eh?
  269. 10:25Because, well, you see it's not cheap,
  270. 10:27but it is true that it's still cheaper.
  271. 10:28But there is another thing that the
  272. 10:30story doesn't tell, and that is speed.
  273. 10:33In fact, I have a few demos here
  274. 10:36because I've had access to Kimi K3 for
  275. 10:38over a week. For over a week. Now I'm
  276. 10:42going to tell you about a "cancel"
  277. 10:43attempt they tried against me the other
  278. 10:45day. Kimi K3 is out, it feels like a
  279. 10:47new DeepSeek moment for the world of...
  280. 10:49this is really something I've been
  281. 10:52testing that I find incredible. I mean,
  282. 10:54it really struck me, I thought, "Wow,
  283. 10:55I'm surprised." Okay, I've been using
  284. 10:57it for a week with early access and
  285. 10:59coding, and it's on par with Claude 3.5
  286. 11:01Opus. I swear it's on par, though now
  287. 11:03I'll tell you some of the fine print so
  288. 11:05you keep it in mind. Sometimes it even
  289. 11:06feels superior, and it's true because,
  290. 11:08on top of that, you can ask it for more
  291. 11:10things regarding cybersecurity and
  292. 11:12vulnerabilities without fear of it
  293. 11:14deflecting. Competing with GPT 5.6, Sol
  294. 11:16, and Fable 5, it costs half as much as
  295. 11:18Opus. It's huge. It will be an open
  296. 11:20model starting July 27th. The price.
  297. 11:22Open models are no longer six months
  298. 11:24behind closed ones. We just aren't
  299. 11:25aware of where we are. Now they compete
  300. 11:27directly, okay? And I posted this. What
  301. 11:29happened? This is the tweet of what it
  302. 11:31told me. Juan Pablo Moreno told me.
  303. 11:33You've supposedly been using it for a
  304. 11:35week with early access and coding.
  305. 11:38Right, because I said I'd been using it
  306. 11:39for a week. There's one thing I don't
  307. 11:40understand. If Kimica 3 came out
  308. 11:42yesterday, why are you posting this
  309. 11:45news a day later? The normal thing
  310. 11:47would be to post what you're posting
  311. 11:49today yesterday, alongside the launch
  312. 11:51of Kimica 3. But Juan Pablo doesn't
  313. 11:53stop there because he says, "Damn, eh,"
  314. 11:55he says, "Well, at least the guy stays
  315. 11:57there and that's it.""But no, he went
  316. 12:00so far as to insinuate that I was lying
  317. 12:03." Okay, he was telling me here, "I'm
  318. 12:05not talking about leaking before, I'm
  319. 12:07talking about sharing after the news
  320. 12:08broke." He delayed 24 hours or more to
  321. 12:11share it. You think, if you're someone
  322. 12:14with that privilege, you take advantage
  323. 12:15and share the news an hour or minutes
  324. 12:17after it breaks. Exactly. It was 21
  325. 12:20hours ago that I posted this tweet.
  326. 12:23Meaning, if what he claims were true,
  327. 12:25you wouldn't wait for the other tweet
  328. 12:27to make the post that he makes. To me,
  329. 12:29that post of his is partly a lie. Let's
  330. 12:32see. Look, this is the official email,
  331. 12:35and just so you see, I'm going to
  332. 12:38reload, uh, from Moonshot AI for Kimi,
  333. 12:41okay? with the NDA signature, and this
  334. 12:44is from July 13th. Regardless, I
  335. 12:48already had access; I had access even
  336. 12:50before signing the NDA, but anyway,
  337. 12:52just so you see it's not a lie, I'll
  338. 12:54refresh, okay? So you can see the email
  339. 12:58appears there, there you have it, okay?
  340. 13:00There you have it. Second, because it's
  341. 13:02impressive. I already had a bunch of
  342. 13:05demos; in fact, we can even see the
  343. 13:07days I used the last demos I made, the
  344. 13:10Black Hole. The Midw Strike, the Neon
  345. 13:13Tetris. Actually, let's take a look at
  346. 13:15Midw Strike, while we're at it. I made
  347. 13:16a Counter Strike, a Midw Strike, two
  348. 13:18prompts, because for the first prompt I
  349. 13:19had to fix an error. First off, it has
  350. 13:21the buy menu and everything, eh? You
  351. 13:22have the buy menu. Let's buy the Desert
  352. 13:25Eagle, okay? Eh, get rid of this. Then
  353. 13:28you have the scoring and everything.
  354. 13:31You can move down, I mean, notice that
  355. 13:33the crosshair works well. Obviously, in
  356. 13:35terms of design, it's quite improvable.
  357. 13:37I'm telling you, I went for it. Watch
  358. 13:39out, watch out, watch out. I didn't get
  359. 13:44him. What a jerk. Well, you can plant
  360. 13:46the bombs and everything. You can plant
  361. 13:48the bombs in two prompts. I mean, what
  362. 13:51surprised me. Oh, what a rat, what a
  363. 13:54rat, what a rat, man. And look at the
  364. 13:57camera. I don't know, I don't know. I
  365. 13:59mean, incredible, incredible,
  366. 14:01incredible. But aim well, man. But he
  367. 14:03really pulled a rat move on me there,
  368. 14:05man. The wretch just camped there. Look
  369. 14:07, here you have all the stuff for the
  370. 14:09menu. I don't know what else. The last
  371. 14:11one can change the Oh, look at him,
  372. 14:12look, look at those rats, how they run,
  373. 14:14eh? Watch out. Oh, oh, wow, he got me
  374. 14:22again. He took it from me. Look, you
  375. 14:24can reload. He took it from me. This is
  376. 14:26one. Let's go with Let's go with others
  377. 14:28. The Black Hole one. Well, the Black
  378. 14:30Hole one too. Look. and he adds sound,
  379. 14:33man. It's really cool, eh? I mean,
  380. 14:36Chimera 3 surprised me because it does
  381. 14:38some very cool stuff, eh? Very cool.
  382. 14:40Well, here you can rotate. It's like a
  383. 14:42black hole, okay? And you can you can
  384. 14:45uh well zoom out and such and you can
  385. 14:48hit warp. Look at this. Let's see,
  386. 14:54what's really surprising is like the
  387. 14:57quality of the things it makes, man.
  388. 15:01And I made a CRUD, things like that,
  389. 15:03but what's really surprising is the
  390. 15:05quality it makes, you know? I mean, you
  391. 15:08see it and you say, "Damn, eh, it's not
  392. 15:10bad at all." Eh, look, here you have a
  393. 15:12fairly easy CRUD. In this case, it's a
  394. 15:14single prompt and it didn't create a
  395. 15:15very good architecture or anything. So
  396. 15:17let's ask it so we can see it in action
  397. 15:19. I'm going to tell it, I'll put it in
  398. 15:21YOLO mode, okay? Chemistry One, here we
  399. 15:24have it, and I'll tell you, uh, it
  400. 15:26improves the architecture to make it
  401. 15:28much more scalable and maintainable
  402. 15:31with best practices, okay? As for
  403. 15:33speaking in English and such, Spanish,
  404. 15:35all that is perfect. What's the problem
  405. 15:38I see with Chemistry 3, at least while
  406. 15:41I've been using it? The little problem
  407. 15:44it has, which I don't know if it will
  408. 15:46improve, is the issue of speed, I mean,
  409. 15:47tokens per second. When you use it
  410. 15:50sometimes you notice it struggles, you
  411. 15:52know? I understand it's for a reason
  412. 15:56they've mentioned themselves, which is
  413. 15:57that they are slammed with the GPU
  414. 15:59issue, I mean, they can't keep up. In
  415. 16:02fact, Chemistry 3 has received more
  416. 16:03love than we expected and our GPUs are
  417. 16:05feeling it. In the last 48 hours,
  418. 16:07demand has been close to the limits of
  419. 16:08our capacity. To protect the experience
  420. 16:11of existing subscribers, we are
  421. 16:12temporarily pausing new subscriptions
  422. 16:15and prioritizing compute for current
  423. 16:17members. Existing subscribed users are
  424. 16:20not affected. And that's why, friends,
  425. 16:23if you try to go and use Chemistry 3
  426. 16:25and subscribe, you're going to find a
  427. 16:28surprise and that is that you can't
  428. 16:30subscribe, it won't let you subscribe.
  429. 16:33So, bad news. The good thing is that
  430. 16:37since it's available in other places,
  431. 16:39for example, you can use it on Open
  432. 16:41Code Go, well, since they've already
  433. 16:43added it there, you could subscribe to
  434. 16:45Open Code Go and directly give it a
  435. 16:47shot there. The problem is that as you
  436. 16:51can see, even though you now have 2x
  437. 16:54usage, you only have 220 requests
  438. 16:56available within the 5-hour limit. Only
  439. 17:00220. The difference compared to the
  440. 17:03rest is quite, quite brutal, huh? As I
  441. 17:06tell you, the problem I see, oh look,
  442. 17:07but why did it? Ah, because it made me
  443. 17:09a plan. Okay, so here we have the plan
  444. 17:11and such. I say, if I put the YOLO,
  445. 17:13what for? Well, here we have the plan.
  446. 17:16Honestly, the code quality is very,
  447. 17:18very surprising, the things it does for
  448. 17:203D stuff, I mean, for making 3D games,
  449. 17:22incredible, huh? When you try it, it's
  450. 17:25going to blow your mind and that's when
  451. 17:27you'll have a real moment. Wow. Here
  452. 17:29you have the good news, that you also
  453. 17:31have it available on Open Gold. The bad
  454. 17:34thing is that, the limits, at least per
  455. 17:36week, for Chemistry 3 are only 250 and
  456. 17:39per month it only gives you 490 calls.
  457. 17:43It's quite low, huh? It runs out
  458. 17:45quickly. So you'll have to keep
  459. 17:47changing models in Open AI or you'll
  460. 17:49run out, and you see that in 5 hours
  461. 17:52you have 220 now temporarily, but soon
  462. 17:55it will only be 110 per week, 250. I
  463. 17:58mean, very, very little. You can really
  464. 18:00tell it's much more expensive, huh?
  465. 18:02Well, there you can see how it's
  466. 18:04already making changes. Let's see if we
  467. 18:06can see the changes here. Okay, we can
  468. 18:08see the tasks a bit now, okay? With the
  469. 18:11configuration, with TypeScript. Look,
  470. 18:14it's added some HTTP modules here to
  471. 18:17separate. Look, model, repository,
  472. 18:20routes. Man, honestly, it's going to
  473. 18:22surprise you a lot, huh? Now, like I
  474. 18:24said, there have been some issues with
  475. 18:26speed. The week I had access, I didn't
  476. 18:29have any problems. For example, Counter
  477. 18:32Strike took quite a while to build, but
  478. 18:34that’s because it generated a lot of
  479. 18:36code, but once they opened it up, wow,
  480. 18:38it’s been tough, really tough. So
  481. 18:41anyway, let's see how it goes. Let's
  482. 18:42see how it goes. I'm very curious to
  483. 18:44see what you think. Let's see what you
  484. 18:46think. That said, I have to tell you a
  485. 18:48few more things about Kimi. First off,
  486. 18:50the founder. No, not the founder, uh,
  487. 18:52maybe co-founder. Uh, Chun Chang. I’m
  488. 18:55not sure if he’s a co-founder or what
  489. 18:56, but anyway. Chuchan says: "Sure, many
  490. 19:00people...Moonshot, in their own blog
  491. 19:02post, mentions this: limitations, that
  492. 19:05it's been trained to keep a history of
  493. 19:08thoughts and that you should be careful
  494. 19:11with the harness you use, and so on,
  495. 19:13and that you should try to avoid
  496. 19:16switching to Kimi 3 in the middle of a
  497. 19:18session, okay?". And then, that it's
  498. 19:22too proactive, and that really is true,
  499. 19:24and it's one of the reasons why you see
  500. 19:26that sometimes it takes longer than it
  501. 19:28should to finish tasks. It's like if it
  502. 19:31finds a small problem, it starts
  503. 19:33overthinking it, and I think that can
  504. 19:34be significantly improved, and like I
  505. 19:36said, it’s still struggling a bit to
  506. 19:38not take so long. But then it says,
  507. 19:42although the model is quite competitive
  508. 19:45in general, there is still a noticeable
  509. 19:48gap in the user experience compared to
  510. 19:51Claude 3.5 and GPT-4o. They are the
  511. 19:54ones saying this, okay? This is the
  512. 19:55Moonshot team in their own blog,
  513. 19:57commenting on this. They themselves
  514. 20:00don't see it at the level of GPT-5.6
  515. 20:02Sol or Claude 3.5 Sonnet. I think
  516. 20:05it’s true that it’s not at that
  517. 20:07level, but it’s not that far off
  518. 20:09either, which is the really surprising
  519. 20:10part. Now, what's going on? Well,
  520. 20:14someone from Moonshot has already
  521. 20:16replied to this tweet and said, "Watch
  522. 20:19out and wait for Kimi 3.1." And they
  523. 20:22just released the third one. Regarding
  524. 20:25the issue about them pausing
  525. 20:26subscriptions and all that, I have to
  526. 20:28comment on this post that says there
  527. 20:30are two types of companies, and I agree
  528. 20:32that when they run out of compute
  529. 20:34capacity for new clients: There are
  530. 20:35companies like Moonshot that stop new
  531. 20:37sign-ups to avoid delays for existing
  532. 20:39customers. And it says Anthropic
  533. 20:42reduces usage for existing customers by
  534. 20:4450%to accommodate new ones. That tells
  535. 20:47you everything you need to know about
  536. 20:49why Anthropic hates open source. I
  537. 20:52don't know if that's the reason why
  538. 20:54they hate open source, but it is true
  539. 20:56that, hey, I respect Moonshot for
  540. 20:58stopping subscriptions as soon as they
  541. 21:00saw they couldn't support new users,
  542. 21:03because honestly, you could tell it
  543. 21:05wasn't running that smoothly. Look, in
  544. 21:08this case it’s finished; I think the
  545. 21:10speed is improving. Uh, well, here we
  546. 21:12can see how they’ve also separated
  547. 21:14the database side of things. Well, it's
  548. 21:16a fairly simple database. In the end,
  549. 21:17it’s a pretty simple CRUD, it’s not
  550. 21:19like there’s much to it. Let’s see,
  551. 21:21here, look, create task repository.
  552. 21:23Okay, sure. And it doesn't execute
  553. 21:25until it gets here. I really like that
  554. 21:27they used functions instead of classes,
  555. 21:29okay? That they made it functional, I
  556. 21:30like it. And here, let's see, the task
  557. 21:33repository, look, it passes it, so it
  558. 21:35has dependency injection here. They
  559. 21:37didn't build a hexagonal architecture,
  560. 21:38by any means, but at least it's true
  561. 21:40that they separated things, and look,
  562. 21:42they added some tests here, so, well,
  563. 21:43looks good. Looks good, honestly. Uh,
  564. 21:45other good things about this. You can
  565. 21:47use it in Cursor. Notice that you can
  566. 21:49use Kimi or others. You can choose it
  567. 21:52right there. Notice that they have it
  568. 21:54selected. Either that or the API, you
  569. 21:56can choose. So you can connect it, you
  570. 21:59can really connect it with Claude, with
  571. 22:01Cursor, with whoever you want. And also
  572. 22:05, also, uh, one thing that is
  573. 22:06interesting, which is what I mentioned
  574. 22:08at the beginning, that it has open
  575. 22:10weights, the good thing is that if
  576. 22:12Chemistry 3 lacks GPUs for whatever
  577. 22:14reason, because they can't handle the
  578. 22:16demand, the good thing is what Chiao Q
  579. 22:18says, and she is the founder of Ticon
  580. 22:20AI, which is a company that is in
  581. 22:22Redwood City. Well, anyway, the beauty
  582. 22:26of open weights is that when Chemistry
  583. 22:283 lacks GPUs, when it doesn't have
  584. 22:29infrastructure, many neocloud clouds
  585. 22:31with GPUs can join in. Kimi will
  586. 22:34release its weights on July 27th, to
  587. 22:36host Kimy K3 and monetize. That is why
  588. 22:38we need a free market. And it's true,
  589. 22:40and that is one of the good things
  590. 22:41about open weights that I don't
  591. 22:43understand why many people say, "No,
  592. 22:44it's just that, if I can't host it on
  593. 22:46my machine, it's useless." No, my
  594. 22:47friend, that's not what it means. I
  595. 22:49mean, what it really means is that what
  596. 22:51you will be able to do is not be tied
  597. 22:53to one company, because in the end,
  598. 22:55using Chemistry 3 you can use it on any
  599. 22:57provider that gives you the access, of
  600. 22:59course, and usually, since there will
  601. 23:01be more competition, it will give you
  602. 23:03more possibilities. Here we have Rohan
  603. 23:06saying that US companies like Modal,
  604. 23:08Fireworks, and Baseten, which are
  605. 23:10companies like Fireworks AI, this is a
  606. 23:12web that gives you artificial
  607. 23:14intelligence inference, okay? And so it
  608. 23:17gives you access to models like, for
  609. 23:19example, Chemistry 2.7, which will be
  610. 23:20Chemistry 3. They give it to you at a
  611. 23:23quite attractive price, but it also
  612. 23:25gives it to you from US servers and at
  613. 23:26a quite good speed, which is that this
  614. 23:28is very, very interesting. They will be
  615. 23:31able to serve Chemistry 3 at a cost 10
  616. 23:33times lower than that of their Chinese
  617. 23:35competitors because they have access to
  618. 23:37advanced Nvidia and AMD chips. Of
  619. 23:39course, it's just that in China it's
  620. 23:41difficult to access Nvidia chips and
  621. 23:43therefore they won't be able to use
  622. 23:45them to serve their own models, which
  623. 23:47I'm sure they use some, okay? But it is
  624. 23:49true that this is a competitive
  625. 23:50advantage that people in the United
  626. 23:52States give you. There is a certain
  627. 23:53irony in this. A large part of the
  628. 23:55research and model development has
  629. 23:57moved to China, but once the model is
  630. 23:58optimized for next-generation hardware,
  631. 24:00like Nvidia's Rubin architecture, its
  632. 24:01operating cost could be reduced by 10
  633. 24:03to 100 times. This is stated by the
  634. 24:05co-founder of Stability AI. It is also
  635. 24:07true that there is one thing that is
  636. 24:08not taken much into account, I don't
  637. 24:10have data, okay? But we would have to
  638. 24:12see what the actual cost of electricity
  639. 24:15is, because it's true that the cost of
  640. 24:17GPUs and such is significant, having
  641. 24:19access to the GPU is important, but
  642. 24:21what is the cost these electricity
  643. 24:24companies in China face? Because I'm
  644. 24:26not very clear on that. At the domestic
  645. 24:28level, I know that electricity in China
  646. 24:29is quite cheap, and I imagine that at
  647. 24:31the...of course, you have to consider
  648. 24:32the cost, you have to take everything
  649. 24:34into account, it’s GPUs. So, I don't
  650. 24:36know if electricity is cost-effective
  651. 24:37for China, because I know that in the
  652. 24:39United States it’s an issue that
  653. 24:40costs them quite a bit more. Sure, in
  654. 24:42China they are installing a ton of
  655. 24:43solar batteries, I think they also have
  656. 24:45quite a bit of nuclear power, so it may
  657. 24:47be that this gives them an advantage in
  658. 24:48this regard that is often not taken
  659. 24:50into account, and I think it is also
  660. 24:52important. If that doesn't seem like
  661. 24:53much because you say, well, that's it,
  662. 24:55okay, Kimi, well, it's not like Qwen
  663. 24:56didn't come out too. Qwen 3.8 also
  664. 24:58arrived this week and it’s also a 2.4
  665. 25:01trillion parameter model, second only
  666. 25:04to Fable 5, according to them. It's
  667. 25:07already available on Alibaba Token Plan
  668. 25:09and it will also be open source, and
  669. 25:11they have already said that they are
  670. 25:12also going to launch more than one
  671. 25:14model of different sizes; I mean, this
  672. 25:16is the largest as well as the most
  673. 25:17advanced, but the idea is that they
  674. 25:19will be launching more, so that open
  675. 25:21models offer you almost the same as the
  676. 25:23best models at half the price. Well,
  677. 25:26that's the good part. Sure, see, this
  678. 25:27is the issue; I don't understand these
  679. 25:29comments. It says, "2.4 trillion open
  680. 25:31parameters is impressive, but almost no
  681. 25:33one is going to serve that monster."
  682. 25:36The real impact of Qwen is not the
  683. 25:37giant model that you download and can't
  684. 25:39run. It's the quantized and small
  685. 25:41versions that do fit on your GPU, open
  686. 25:43models you can't run are only open in
  687. 25:44name. You can't reduce the world to
  688. 25:46what you can run on your GPU. I
  689. 25:48understand what they mean, but I don't
  690. 25:50understand where it’s coming from. I
  691. 25:51just gave an example. Fireworks AI, for
  692. 25:54example: the fact that you can't run it
  693. 25:56on your machine doesn't mean that an
  694. 25:58infrastructure company, like Fireworks
  695. 26:00for example, can't come along and offer
  696. 26:02you, for example, look, Qwen 3.7 Plus,
  697. 26:05and they can indeed resell it to you or
  698. 26:07host it for you and let you have it at
  699. 26:09a very attractive price. They can offer
  700. 26:13it to you at a very attractive price,
  701. 26:15and of course, if they give you Qwen
  702. 26:173.8 Och, where you can also run it from
  703. 26:20different providers, you're not tied to
  704. 26:22a closed provider, you know? Like it
  705. 26:25doesn't just have to be on OpenAI. The
  706. 26:27fact that it's open means that not only
  707. 26:29can one person launch it, but you'll be
  708. 26:31able to launch it in different places,
  709. 26:32and on top of that, you'll be able to
  710. 26:34have it available geographically in
  711. 26:36different locations, which is more
  712. 26:38interesting. I don't know, I mean, I
  713. 26:40don't understand, I understand what
  714. 26:41they mean, but I don't get where it's
  715. 26:42coming from. Because I didn't even say
  716. 26:44in my post, "You can install it on your
  717. 26:46machine." No, I don't say anything like
  718. 26:48that. And yet, many people often reduce
  719. 26:51it to, "If I can't run it on my machine
  720. 26:52, it's no use to me." But look, they
  721. 26:55are open weights, and the beauty of it
  722. 26:57is that we can have it in more than one
  723. 26:59place. I don't know, I don't know, it's
  724. 27:01a bit of a weird comment because I see
  725. 27:03it repeated a lot. Well, if you don't
  726. 27:05have enough with Qwen 3.8, it's also
  727. 27:07coming soon and, well, there are very
  728. 27:09strong rumors. GLM 5.5, they're going
  729. 27:13to skip GLM 5.3, and GLM 5.5 is coming,
  730. 27:15which would launch in August, which
  731. 27:18would be over a billion parameters,
  732. 27:21also getting closer, and well, context
  733. 27:23as well. So, there you go, you can see,
  734. 27:25my goodness, how things are, how things
  735. 27:27are going. With all this, I'm going to
  736. 27:30tell you what my favorite models are. I
  737. 27:32haven't tested Qwen yet, I haven't
  738. 27:34tested it yet, so it doesn't count, but
  739. 27:36this week, among all the models I've
  740. 27:38been using, these are my favorite
  741. 27:40models. Simple tasks, GPT 5.6 Luna Max.
  742. 27:44Daily coding, GPT 5.6 Sol Medium. I
  743. 27:48know this seems very geeky, the whole "
  744. 27:50Max,""Medium," and so on, but all this
  745. 27:52has been studied, and in fact, today
  746. 27:54when I finish the stream, I'm going to
  747. 27:56post a video on YouTube where I test
  748. 27:58GPT 5.6 and explain why. I'm explaining
  749. 28:01this so you can understand it. Then,
  750. 28:03for complex problems, GPT 5.6 Sol Max;
  751. 28:06for quality and speed, Grok 4.5. Grok
  752. 28:114.5 has surprised me a lot, I mean,
  753. 28:12I've been seriously blown away by Grok,
  754. 28:14especially using it with Grok Build.
  755. 28:16I've been blown away. I have to test it
  756. 28:18more, but it has surprised me a ton. Eh
  757. 28:20, I don't know if it's the model, I
  758. 28:21don't know if it's the data centers, I
  759. 28:22have no idea, but it seems spectacular
  760. 28:24to me. For interfaces, uh, UI, Fable 5,
  761. 28:27vulnerabilities, and security. Another
  762. 28:29call-out I've received, another
  763. 28:30call-out. Vulnerabilities and security.
  764. 28:33A lot of people in many places,
  765. 28:35everywhere at once, accused me of lying
  766. 28:37, saying it was bait. They told me, "
  767. 28:40You can't use Fable 5 for
  768. 28:42vulnerabilities and security, liar, I
  769. 28:44don't know what.""It's so obvious that
  770. 28:46the list was written by, uh, ChatGPT,
  771. 28:49because working on cybersecurity in
  772. 28:51Fable redirects you to Opus 4.8," they
  773. 28:54tell me. And this isn't true. I love
  774. 28:57you, Alexis, because I know you don't
  775. 28:58say it with bad intentions. Look, you
  776. 29:00can try it yourselves because I use it
  777. 29:02for that a lot. Audit the security of
  778. 29:04my endpoints and make sure I don't have
  779. 29:05vulnerabilities. And it will do it
  780. 29:07perfectly for you. It will do it
  781. 29:08perfectly for you. I mean, you won't
  782. 29:10have any problem at all. It won't kick
  783. 29:11you out of Fable 5. It will do it from
  784. 29:13start to finish, without any issues.
  785. 29:15You won't have any problem, like, it
  786. 29:17just won't tell you anything. Now, if
  787. 29:20you ask it something like this, "How
  788. 29:22can I attack this website to exploit
  789. 29:24its vulnerabilities?" Then yes, it will
  790. 29:27kick you out. Okay, it will kick you
  791. 29:30out, but try it yourselves with Fable 5
  792. 29:32, uh, any API you have, you ask it, "
  793. 29:34Review these vulnerabilities, improve
  794. 29:37the defense of this, prevent them from
  795. 29:39doing this," and so on, it doesn't have
  796. 29:41any problem. There's no problem at all.
  797. 29:43And it doesn't just, it doesn't have
  798. 29:45any problem, it's super good. Uh, of
  799. 29:46course, if you say, "I'm a hacker, I'm
  800. 29:48a malicious hacker and what I want is
  801. 29:50for you to find a vulnerability here to
  802. 29:52attack this bank.""Yes, it kicks you
  803. 29:54out." But I know many people got stuck
  804. 29:55on that, they haven't tried it again,
  805. 29:57but it's not that, really, I mean, I
  806. 29:59use it a ton for this and it actually
  807. 30:01does it very well. But look, it's not
  808. 30:03just Alexis who says it to me, because
  809. 30:05Alexis, for example, I know he doesn't
  810. 30:06say it in a mean way, I know he doesn't
  811. 30:08say it to be mean, but there are others
  812. 30:09who did say it to me a bit. "It's bait,
  813. 30:11the Fable post, it can't be used for
  814. 30:13anything related to vulnerabilities and
  815. 30:16security." But why are you writing to
  816. 30:18me if you haven't even tried it? You
  817. 30:19haven't tried it. You read it in an
  818. 30:21article, you read it somewhere out
  819. 30:23there. You don't know what you're
  820. 30:25talking about, man. I mean, the Bite is
  821. 30:26your twist, jerk, because you don't
  822. 30:28know what you're talking about. You
  823. 30:29haven't even tried it. I don't know,
  824. 30:31maybe he has tried it. And it's true,
  825. 30:33it's true that, especially at the
  826. 30:34beginning, the first day, the first day
  827. 30:36they opened it, it's true that it was
  828. 30:38quite a bit more nitpicky, I mean...But
  829. 30:39I've used it a ton. But anyway, I don't
  830. 30:42know, I don't know, I don't know,
  831. 30:43people don't believe me. Now, it is
  832. 30:45true that I have a friend who works in
  833. 30:47chemistry and this is true, she tells
  834. 30:49me that regarding her work topics, you
  835. 30:51absolutely cannot ask her anything. Uh,
  836. 30:54about chemistry topics, that she does
  837. 30:56say is impossible, you can ask her
  838. 30:58however you want, there's no way,
  839. 30:59that's true. But look, regarding
  840. 31:01vulnerabilities, I'm telling you
  841. 31:03there's no problem at all. You ask it
  842. 31:05for an audit, improvements, whatever
  843. 31:07for defense, no problem. At least in my
  844. 31:09case, and I'm telling you I'm using it
  845. 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.