YouTube2Text

Kimi K3 la está ROMPIENDO: El modelo chino que compite con GPT-5.6 y Claude Fable 5 — Transcript

by Fazt · 2,145 words · 305 segments · language en · Watch on YouTube

Full transcript

  1. 0:00A few hours ago, Kimi 3 was released,
  2. 0:02which is the most updated version of
  3. 0:04the Kimi model, this multimodal model
  4. 0:06that allows generating text, images,
  5. 0:08videos and also code. Well, today there
  6. 0:10has been a lot of talk about what it
  7. 0:12can generate, which surprisingly is
  8. 0:13jumping quite a bit in the benchmarks
  9. 0:15to the point that it is reaching Cloud
  10. 0:17Fable 5 or even GPT 5.6 Sol. Well, to
  11. 0:19do a demo today I am going to show you
  12. 0:21some examples, in addition to also
  13. 0:22being able to summarize what this new
  14. 0:24model is about. Let's get started.
  15. 0:26Coders, when an application is already
  16. 0:28in production, one of the biggest
  17. 0:30challenges is detecting problems before
  18. 0:32users find them. For example, a page
  19. 0:34that stops loading, an API that starts
  20. 0:36responding slowly or a flow within the
  21. 0:38application that simply stops working.
  22. 0:40And waiting for users to report these
  23. 0:42problems is not always the best
  24. 0:44strategy. This is where synthetic
  25. 0:46monitoring comes in. This technique
  26. 0:48consists of simulating interactions
  27. 0:49with your application automatically, as
  28. 0:51if they were real users. For example,
  29. 0:53you can set up tests that periodically
  30. 0:55check whether a page loads correctly,
  31. 0:57whether a form works, or whether an API
  32. 0:59is responding as it should. And these
  33. 1:01tests are run from different locations
  34. 1:03and constantly, allowing you to detect
  35. 1:06availability or performance issues
  36. 1:07before they affect real users. And to
  37. 1:10solve this, you can use Open Manager
  38. 1:12Nexus, where you can set up this type
  39. 1:14of monitoring for web pages, APIs, and
  40. 1:16entire applications. In addition, when
  41. 1:18a problem occurs, you can receive
  42. 1:20real-time alerts, which helps you react
  43. 1:22quickly. And the great thing is that
  44. 1:24you can deploy Open Manager Nexus both
  45. 1:25in the cloud and on your own
  46. 1:27infrastructure. In addition, right now
  47. 1:29you have a free 30-day trial available
  48. 1:30on their official website. So take
  49. 1:32advantage of this opportunity and try
  50. 1:34out all the features using the link
  51. 1:35I've left in the description. Alright,
  52. 1:37let's start by reading a little bit
  53. 1:39about what the announcement would be
  54. 1:40about this new Kimi model called K3,
  55. 1:42which is basically the successor to K
  56. 1:442.7. Now, what has attracted the most
  57. 1:46attention among all the things or why
  58. 1:47this model is being mentioned so much
  59. 1:49on the internet is mainly because it is
  60. 1:50now competing with the larger models.
  61. 1:52That is to say, in Chemistry 2.7 we
  62. 1:54compared that model with GPT 5.3 or
  63. 1:57Opus 4.7. So, when Chemistry 3 comes
  64. 1:59out, one would expect it to compete
  65. 2:01with Opus 4.8 or GPT 5.5, but in fact
  66. 2:03it is not. And it is because it has
  67. 2:05made a huge leap and now it is being
  68. 2:07compared with GPT 5.6 Sol and Fable 5,
  69. 2:09which are the largest and most produced
  70. 2:11models available at the moment. Which
  71. 2:15means that the 6-month gap in which the
  72. 2:16Chinese models were behind the most
  73. 2:18popular models is actually not that big
  74. 2:19. And at this point it is quite
  75. 2:22surprising that a model that was almost
  76. 2:24in the last lists, outside the 10th
  77. 2:26ranking of many benchmarks, is now
  78. 2:28competing with these two models here,
  79. 2:30almost similar to what happened with
  80. 2:32GLM 5.2 a few weeks ago. But well, to
  81. 2:34summarize, this is a new model that
  82. 2:36reaches 3 trillion parameters and, in
  83. 2:37order to, let's say, increase its speed
  84. 2:39a little, especially since it is a
  85. 2:40large model, they have a new way of
  86. 2:42implementing what would be the decoding
  87. 2:43of tokens at the time of input, and it
  88. 2:45is precisely with these new methods
  89. 2:47that it is called Delta Attention and
  90. 2:48Attention Residuals. What this means in
  91. 2:51practice is that we also have a million
  92. 2:53tokens in a context window and that
  93. 2:55this allows us to apply it to long-term
  94. 2:57tasks such as executing assistants or
  95. 2:59executing code or reasoning and
  96. 3:01basically agent tasks. Although well,
  97. 3:03one way of measuring all this is using
  98. 3:05a benchmark that does not come from the
  99. 3:07same company. So, we also have
  100. 3:09Artificial Analysis, which is a
  101. 3:10benchmark that measures multiple models
  102. 3:11in different types of areas. And, for
  103. 3:13example, when we talk about
  104. 3:15intelligence, especially when writing
  105. 3:17code or performing agent tasks, what
  106. 3:18has happened with this model, simply if
  107. 3:21we compare it with one of its
  108. 3:22predecessors, which is 2.6, for example
  109. 3:24, has made a leap almost on par with
  110. 3:26what it is now with GPT 5.6 and Fable.
  111. 3:28And well, this applies not only when
  112. 3:30generating code, but also when it is
  113. 3:32utilized in tasks to execute programs
  114. 3:34or, once again, for those tasks that
  115. 3:36are related to agents in a general
  116. 3:37sense. Although, well, something very
  117. 3:39important with all this is precisely to
  118. 3:41know how expensive it is and in fact it
  119. 3:43is not such an expensive model to be a
  120. 3:44model that is competing with those
  121. 3:46large models. Obviously the answer may
  122. 3:48vary, but just to give us an idea and
  123. 3:49to measure the cost. You can see here
  124. 3:51that this is a little more expensive
  125. 3:53than what GLM 5.2 would be, but it is
  126. 3:55still cheaper than Opus 4.8, for
  127. 3:57example. And well, it is also much
  128. 3:59cheaper than GPT. Although, well,
  129. 4:01another very important metric regarding
  130. 4:02the model is the response speed. And
  131. 4:04well, it was already known that this
  132. 4:05model was not very fast. In fact, if
  133. 4:07you have used K 2.7, you know that it
  134. 4:09is a model that can be quite decent,
  135. 4:10but in terms of speed, there are much
  136. 4:12cheaper models. Even a descendant of
  137. 4:14this, which could be Composer 2.5, is
  138. 4:16faster in my opinion and it shows. And
  139. 4:18in this case, it has not improved that
  140. 4:20much in this aspect, so at least in
  141. 4:22terms of speed, do not expect a huge
  142. 4:23change. But hey, in the end, what
  143. 4:25matters is the code generation and how
  144. 4:26good it is. So let's see some examples.
  145. 4:28And well, to be able to try it, you
  146. 4:30already have it available in what would
  147. 4:32be your Kimi.com subscription or you
  148. 4:33can use it in applications such as Kimi
  149. 4:35Word, Kimi Code or Kimi API. And well,
  150. 4:37just to give you an exact number, in
  151. 4:39this case according to their own
  152. 4:41documentation, the entry million tokens
  153. 4:42cost around $ 3 and the exit costs
  154. 4:44around 15. Again, placing it far behind
  155. 4:46what would be 5.5 and 4.8 and obviously
  156. 4:48far behind what would be Fable 5. In
  157. 4:50terms of prices we are talking. And
  158. 4:52well, to start using it, in my case I
  159. 4:54am going to be testing it from Kimi,
  160. 4:55from Kimi Code, their tool or harness.
  161. 4:57And well, from here you can start
  162. 4:59changing the model, simply type slash
  163. 5:00model and with this you will start to
  164. 5:02see these options. We now have the K3
  165. 5:03option. And by the way, if you want to
  166. 5:05know how to use Kimi Code or want to
  167. 5:07create some examples, I have already
  168. 5:08created videos previously using this
  169. 5:09harness, so I will also leave it in the
  170. 5:11description. Okay, let's ask for a
  171. 5:13somewhat extensive application and I'm
  172. 5:15going to paste a large text here and,
  173. 5:17basically, I'm going to ask for a
  174. 5:18personal finance dashboard that allows
  175. 5:20us to manage different accounts, set
  176. 5:22financial goals and, basically, keep a
  177. 5:23record of the income, withdrawals and
  178. 5:25other operations that can be done from
  179. 5:27an account. So, let's see how long it
  180. 5:29can take for a project of this style.
  181. 5:31It's not a plan, but it's not small
  182. 5:33either. And well, by the way, if you
  183. 5:35also want me to simply implement it,
  184. 5:37you can write a yolo here and with that
  185. 5:39it will simply start answering the
  186. 5:40questions by itself. Very well, after
  187. 5:43approximately 40 minutes or so, the
  188. 5:44project generation has passed and in
  189. 5:46theory the reason why it has taken so
  190. 5:48long is that I have implemented each of
  191. 5:50the modules and also what would be the
  192. 5:51backend and frontend. And well, what it
  193. 5:54has generated is this here, which
  194. 5:55actually doesn't look bad at all. It
  195. 5:57has already placed demo data, we
  196. 5:59already have sections here to be able
  197. 6:00to have even graphs and well, in terms
  198. 6:02of UI it looks pretty good. Now, in
  199. 6:04this case, this is a model without any
  200. 6:06additional skills, that is, this is the
  201. 6:08model as it responds. And well, this is
  202. 6:10one of the things that is also being
  203. 6:12highlighted quite a bit on social media
  204. 6:14and that is that in the Front Code
  205. 6:16Arena benchmark, what has happened is
  206. 6:17that Kimica 3 is much better than Fable
  207. 6:195 and 5.6 sol. That means that for
  208. 6:22frontend tasks it really surpasses
  209. 6:23those models and it is really
  210. 6:24noticeable in the first instance,
  211. 6:25although again this has taken a long
  212. 6:26time to generate. And another thing
  213. 6:29that I don't see here is precisely the
  214. 6:31way to add data. Apparently it was made
  215. 6:34simply as a frontend to be able to
  216. 6:35visualize things and there is no
  217. 6:37apparent way to be able to add
  218. 6:38additional data. And well, in terms of
  219. 6:41the code, what I want to generate is a
  220. 6:42Next application where the API is,
  221. 6:44there are the frontend applications and
  222. 6:45, in terms of the database, since it
  223. 6:46has not been very explicit about which
  224. 6:48database it should use, it seems that
  225. 6:49here it has chosen SQLite. And well, in
  226. 6:51reality it has not added many packages;
  227. 6:53is basically Nex in its latest versions
  228. 6:55, which is pretty good. That is to say,
  229. 6:57this model already considers at least
  230. 6:58modern versions of these frameworks.
  231. 7:00The next thing that is also included
  232. 7:01are the packages for linting and to
  233. 7:03obviously be able to use TypeScript and
  234. 7:04well, they are not old versions or
  235. 7:06anything like that, so it is quite
  236. 7:07decent to actually produce code at
  237. 7:09least for frontend applications. And
  238. 7:10well, to do a second example, let's ask
  239. 7:12it to create a 3D game in the style of
  240. 7:14Mario Kart, for example, only we are
  241. 7:15going to tell it that it must have
  242. 7:16speed metrics and a realistic scenario.
  243. 7:18Let's see what it creates. And well, if
  244. 7:21we wait about 30 minutes or even a
  245. 7:22little more, we will see that it has
  246. 7:24generated another project here and we
  247. 7:25will check it. In fact, it has given me
  248. 7:28a URL here and this is what it has
  249. 7:29generated. Basically it is a fairly
  250. 7:31basic project using 3GS, obviously, but
  251. 7:33this is a model that is obviously much
  252. 7:35cheaper than GPT 5.6. Yeah, and well,
  253. 7:38we could also do a comparison there one
  254. 7:40of these days, precisely, with the
  255. 7:42other models, but it has even fulfilled
  256. 7:44the idea of the initial pron, which is
  257. 7:45precisely to see in the upper left part
  258. 7:47what would be the time, the speed, and
  259. 7:49on the right side there is like,
  260. 7:51precisely, a map and it is quite good.
  261. 7:53In fact, it is a first pron, I simply
  262. 7:55have not made any more modifications.
  263. 7:57It has taken a while, yes, although
  264. 7:58something else that you can also notice
  265. 8:00is that I am using the maximum effort.
  266. 8:02In fact, if we come here and I put
  267. 8:03slash effort and I press enter, you
  268. 8:05will see that I only have the max mode
  269. 8:06available. So, this means that as the
  270. 8:08variations of this model come out, we
  271. 8:10could get models that respond faster or
  272. 8:12perhaps they give us one or another
  273. 8:14additional variation. But anyway, for
  274. 8:15now it is what we have and again it is
  275. 8:17not bad at all. Even for frontend tasks
  276. 8:19, this model is quite good. And in my
  277. 8:21opinion, I have not tested it that much
  278. 8:23with logic; I mean, I haven't seen what
  279. 8:24it generated. It doesn't look bad at
  280. 8:26all, actually, but I'm still going to
  281. 8:28wait a few days to see some examples
  282. 8:30and, above all, also bring you an
  283. 8:31update on what would be Chemistry 3.
  284. 8:33Well, that's basically today's video.
  285. 8:35The model is quite impressive,
  286. 8:36especially for front-end tasks. I
  287. 8:37really liked it. I haven't been able to
  288. 8:39test it yet, even with parallel tasks
  289. 8:41or with their subagents, but I'm going
  290. 8:42to try it out over the course of this
  291. 8:44week. Well, this just came out, so you
  292. 8:46can also try it out using Kimy Code or
  293. 8:48subscriptions to other harnesses that
  294. 8:49may be out there. Well, if you have any
  295. 8:51questions or suggestions, you can also
  296. 8:52leave them in the comments. See you in
  297. 8:54the next video. And that's all for
  298. 8:56today's video. If you want to know more
  299. 8:58, I'll leave my social links on the
  300. 9:00screen and my website fast.def where
  301. 9:01you can book personalized advice on any
  302. 9:03topic. And don't forget to leave me a
  303. 9:05comment, whether it's a question or a
  304. 9:07suggestion for the next video. See you
  305. 9:08in the next one.

About this transcript

This page contains the full transcript of Kimi K3 la está ROMPIENDO: El modelo chino que compite con GPT-5.6 y Claude Fable 5 by Fazt, generated from the public captions YouTube serves with the video. The transcript has 2,145 words across 305 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.