Kimi K3 la está ROMPIENDO: El modelo chino que compite con GPT-5.6 y Claude Fable 5 — Transcript
Full transcript
- 0:00A few hours ago, Kimi 3 was released,
- 0:02which is the most updated version of
- 0:04the Kimi model, this multimodal model
- 0:06that allows generating text, images,
- 0:08videos and also code. Well, today there
- 0:10has been a lot of talk about what it
- 0:12can generate, which surprisingly is
- 0:13jumping quite a bit in the benchmarks
- 0:15to the point that it is reaching Cloud
- 0:17Fable 5 or even GPT 5.6 Sol. Well, to
- 0:19do a demo today I am going to show you
- 0:21some examples, in addition to also
- 0:22being able to summarize what this new
- 0:24model is about. Let's get started.
- 0:26Coders, when an application is already
- 0:28in production, one of the biggest
- 0:30challenges is detecting problems before
- 0:32users find them. For example, a page
- 0:34that stops loading, an API that starts
- 0:36responding slowly or a flow within the
- 0:38application that simply stops working.
- 0:40And waiting for users to report these
- 0:42problems is not always the best
- 0:44strategy. This is where synthetic
- 0:46monitoring comes in. This technique
- 0:48consists of simulating interactions
- 0:49with your application automatically, as
- 0:51if they were real users. For example,
- 0:53you can set up tests that periodically
- 0:55check whether a page loads correctly,
- 0:57whether a form works, or whether an API
- 0:59is responding as it should. And these
- 1:01tests are run from different locations
- 1:03and constantly, allowing you to detect
- 1:06availability or performance issues
- 1:07before they affect real users. And to
- 1:10solve this, you can use Open Manager
- 1:12Nexus, where you can set up this type
- 1:14of monitoring for web pages, APIs, and
- 1:16entire applications. In addition, when
- 1:18a problem occurs, you can receive
- 1:20real-time alerts, which helps you react
- 1:22quickly. And the great thing is that
- 1:24you can deploy Open Manager Nexus both
- 1:25in the cloud and on your own
- 1:27infrastructure. In addition, right now
- 1:29you have a free 30-day trial available
- 1:30on their official website. So take
- 1:32advantage of this opportunity and try
- 1:34out all the features using the link
- 1:35I've left in the description. Alright,
- 1:37let's start by reading a little bit
- 1:39about what the announcement would be
- 1:40about this new Kimi model called K3,
- 1:42which is basically the successor to K
- 1:442.7. Now, what has attracted the most
- 1:46attention among all the things or why
- 1:47this model is being mentioned so much
- 1:49on the internet is mainly because it is
- 1:50now competing with the larger models.
- 1:52That is to say, in Chemistry 2.7 we
- 1:54compared that model with GPT 5.3 or
- 1:57Opus 4.7. So, when Chemistry 3 comes
- 1:59out, one would expect it to compete
- 2:01with Opus 4.8 or GPT 5.5, but in fact
- 2:03it is not. And it is because it has
- 2:05made a huge leap and now it is being
- 2:07compared with GPT 5.6 Sol and Fable 5,
- 2:09which are the largest and most produced
- 2:11models available at the moment. Which
- 2:15means that the 6-month gap in which the
- 2:16Chinese models were behind the most
- 2:18popular models is actually not that big
- 2:19. And at this point it is quite
- 2:22surprising that a model that was almost
- 2:24in the last lists, outside the 10th
- 2:26ranking of many benchmarks, is now
- 2:28competing with these two models here,
- 2:30almost similar to what happened with
- 2:32GLM 5.2 a few weeks ago. But well, to
- 2:34summarize, this is a new model that
- 2:36reaches 3 trillion parameters and, in
- 2:37order to, let's say, increase its speed
- 2:39a little, especially since it is a
- 2:40large model, they have a new way of
- 2:42implementing what would be the decoding
- 2:43of tokens at the time of input, and it
- 2:45is precisely with these new methods
- 2:47that it is called Delta Attention and
- 2:48Attention Residuals. What this means in
- 2:51practice is that we also have a million
- 2:53tokens in a context window and that
- 2:55this allows us to apply it to long-term
- 2:57tasks such as executing assistants or
- 2:59executing code or reasoning and
- 3:01basically agent tasks. Although well,
- 3:03one way of measuring all this is using
- 3:05a benchmark that does not come from the
- 3:07same company. So, we also have
- 3:09Artificial Analysis, which is a
- 3:10benchmark that measures multiple models
- 3:11in different types of areas. And, for
- 3:13example, when we talk about
- 3:15intelligence, especially when writing
- 3:17code or performing agent tasks, what
- 3:18has happened with this model, simply if
- 3:21we compare it with one of its
- 3:22predecessors, which is 2.6, for example
- 3:24, has made a leap almost on par with
- 3:26what it is now with GPT 5.6 and Fable.
- 3:28And well, this applies not only when
- 3:30generating code, but also when it is
- 3:32utilized in tasks to execute programs
- 3:34or, once again, for those tasks that
- 3:36are related to agents in a general
- 3:37sense. Although, well, something very
- 3:39important with all this is precisely to
- 3:41know how expensive it is and in fact it
- 3:43is not such an expensive model to be a
- 3:44model that is competing with those
- 3:46large models. Obviously the answer may
- 3:48vary, but just to give us an idea and
- 3:49to measure the cost. You can see here
- 3:51that this is a little more expensive
- 3:53than what GLM 5.2 would be, but it is
- 3:55still cheaper than Opus 4.8, for
- 3:57example. And well, it is also much
- 3:59cheaper than GPT. Although, well,
- 4:01another very important metric regarding
- 4:02the model is the response speed. And
- 4:04well, it was already known that this
- 4:05model was not very fast. In fact, if
- 4:07you have used K 2.7, you know that it
- 4:09is a model that can be quite decent,
- 4:10but in terms of speed, there are much
- 4:12cheaper models. Even a descendant of
- 4:14this, which could be Composer 2.5, is
- 4:16faster in my opinion and it shows. And
- 4:18in this case, it has not improved that
- 4:20much in this aspect, so at least in
- 4:22terms of speed, do not expect a huge
- 4:23change. But hey, in the end, what
- 4:25matters is the code generation and how
- 4:26good it is. So let's see some examples.
- 4:28And well, to be able to try it, you
- 4:30already have it available in what would
- 4:32be your Kimi.com subscription or you
- 4:33can use it in applications such as Kimi
- 4:35Word, Kimi Code or Kimi API. And well,
- 4:37just to give you an exact number, in
- 4:39this case according to their own
- 4:41documentation, the entry million tokens
- 4:42cost around $ 3 and the exit costs
- 4:44around 15. Again, placing it far behind
- 4:46what would be 5.5 and 4.8 and obviously
- 4:48far behind what would be Fable 5. In
- 4:50terms of prices we are talking. And
- 4:52well, to start using it, in my case I
- 4:54am going to be testing it from Kimi,
- 4:55from Kimi Code, their tool or harness.
- 4:57And well, from here you can start
- 4:59changing the model, simply type slash
- 5:00model and with this you will start to
- 5:02see these options. We now have the K3
- 5:03option. And by the way, if you want to
- 5:05know how to use Kimi Code or want to
- 5:07create some examples, I have already
- 5:08created videos previously using this
- 5:09harness, so I will also leave it in the
- 5:11description. Okay, let's ask for a
- 5:13somewhat extensive application and I'm
- 5:15going to paste a large text here and,
- 5:17basically, I'm going to ask for a
- 5:18personal finance dashboard that allows
- 5:20us to manage different accounts, set
- 5:22financial goals and, basically, keep a
- 5:23record of the income, withdrawals and
- 5:25other operations that can be done from
- 5:27an account. So, let's see how long it
- 5:29can take for a project of this style.
- 5:31It's not a plan, but it's not small
- 5:33either. And well, by the way, if you
- 5:35also want me to simply implement it,
- 5:37you can write a yolo here and with that
- 5:39it will simply start answering the
- 5:40questions by itself. Very well, after
- 5:43approximately 40 minutes or so, the
- 5:44project generation has passed and in
- 5:46theory the reason why it has taken so
- 5:48long is that I have implemented each of
- 5:50the modules and also what would be the
- 5:51backend and frontend. And well, what it
- 5:54has generated is this here, which
- 5:55actually doesn't look bad at all. It
- 5:57has already placed demo data, we
- 5:59already have sections here to be able
- 6:00to have even graphs and well, in terms
- 6:02of UI it looks pretty good. Now, in
- 6:04this case, this is a model without any
- 6:06additional skills, that is, this is the
- 6:08model as it responds. And well, this is
- 6:10one of the things that is also being
- 6:12highlighted quite a bit on social media
- 6:14and that is that in the Front Code
- 6:16Arena benchmark, what has happened is
- 6:17that Kimica 3 is much better than Fable
- 6:195 and 5.6 sol. That means that for
- 6:22frontend tasks it really surpasses
- 6:23those models and it is really
- 6:24noticeable in the first instance,
- 6:25although again this has taken a long
- 6:26time to generate. And another thing
- 6:29that I don't see here is precisely the
- 6:31way to add data. Apparently it was made
- 6:34simply as a frontend to be able to
- 6:35visualize things and there is no
- 6:37apparent way to be able to add
- 6:38additional data. And well, in terms of
- 6:41the code, what I want to generate is a
- 6:42Next application where the API is,
- 6:44there are the frontend applications and
- 6:45, in terms of the database, since it
- 6:46has not been very explicit about which
- 6:48database it should use, it seems that
- 6:49here it has chosen SQLite. And well, in
- 6:51reality it has not added many packages;
- 6:53is basically Nex in its latest versions
- 6:55, which is pretty good. That is to say,
- 6:57this model already considers at least
- 6:58modern versions of these frameworks.
- 7:00The next thing that is also included
- 7:01are the packages for linting and to
- 7:03obviously be able to use TypeScript and
- 7:04well, they are not old versions or
- 7:06anything like that, so it is quite
- 7:07decent to actually produce code at
- 7:09least for frontend applications. And
- 7:10well, to do a second example, let's ask
- 7:12it to create a 3D game in the style of
- 7:14Mario Kart, for example, only we are
- 7:15going to tell it that it must have
- 7:16speed metrics and a realistic scenario.
- 7:18Let's see what it creates. And well, if
- 7:21we wait about 30 minutes or even a
- 7:22little more, we will see that it has
- 7:24generated another project here and we
- 7:25will check it. In fact, it has given me
- 7:28a URL here and this is what it has
- 7:29generated. Basically it is a fairly
- 7:31basic project using 3GS, obviously, but
- 7:33this is a model that is obviously much
- 7:35cheaper than GPT 5.6. Yeah, and well,
- 7:38we could also do a comparison there one
- 7:40of these days, precisely, with the
- 7:42other models, but it has even fulfilled
- 7:44the idea of the initial pron, which is
- 7:45precisely to see in the upper left part
- 7:47what would be the time, the speed, and
- 7:49on the right side there is like,
- 7:51precisely, a map and it is quite good.
- 7:53In fact, it is a first pron, I simply
- 7:55have not made any more modifications.
- 7:57It has taken a while, yes, although
- 7:58something else that you can also notice
- 8:00is that I am using the maximum effort.
- 8:02In fact, if we come here and I put
- 8:03slash effort and I press enter, you
- 8:05will see that I only have the max mode
- 8:06available. So, this means that as the
- 8:08variations of this model come out, we
- 8:10could get models that respond faster or
- 8:12perhaps they give us one or another
- 8:14additional variation. But anyway, for
- 8:15now it is what we have and again it is
- 8:17not bad at all. Even for frontend tasks
- 8:19, this model is quite good. And in my
- 8:21opinion, I have not tested it that much
- 8:23with logic; I mean, I haven't seen what
- 8:24it generated. It doesn't look bad at
- 8:26all, actually, but I'm still going to
- 8:28wait a few days to see some examples
- 8:30and, above all, also bring you an
- 8:31update on what would be Chemistry 3.
- 8:33Well, that's basically today's video.
- 8:35The model is quite impressive,
- 8:36especially for front-end tasks. I
- 8:37really liked it. I haven't been able to
- 8:39test it yet, even with parallel tasks
- 8:41or with their subagents, but I'm going
- 8:42to try it out over the course of this
- 8:44week. Well, this just came out, so you
- 8:46can also try it out using Kimy Code or
- 8:48subscriptions to other harnesses that
- 8:49may be out there. Well, if you have any
- 8:51questions or suggestions, you can also
- 8:52leave them in the comments. See you in
- 8:54the next video. And that's all for
- 8:56today's video. If you want to know more
- 8:58, I'll leave my social links on the
- 9:00screen and my website fast.def where
- 9:01you can book personalized advice on any
- 9:03topic. And don't forget to leave me a
- 9:05comment, whether it's a question or a
- 9:07suggestion for the next video. See you
- 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.