5 Çin Modeliyle Krallık Kurdum: Sıfırdan Kurulum — Transcript
Full transcript
- 0:00This time, I’ve built a kingdom of
- 0:02models, and inside each of these castles
- 0:03you see, a Chinese model resides.
- 0:06These models run 24/7, all designed to
- 0:08work with each other, and in this video
- 0:10, I will explain how I did it from start
- 0:13to finish and even share their
- 0:14performance with you.
- 0:16If you remember, I had previously
- 0:18created an agent company. I had also set
- 0:20up an agent office consisting of cats.
- 0:23Now, I wanted to add an extra kingdom
- 0:25consisting solely of Chinese models to
- 0:27that. I do most of my work with Cloud
- 0:28Code, but the limits of the most
- 0:29powerful models aren't enough for every
- 0:31task. We agree on this. For tasks like
- 0:33writing cards, fixing text, or setting
- 0:35up a simple page, we don't always need a
- 0:37powerful model. I don’t want to spend
- 0:39money on the most expensive model for
- 0:40these. Because I will do the truly
- 0:42difficult tasks locally in Cloud Code.
- 0:43But I will set up the architecture for
- 0:45my Chinese models on a server. This way
- 0:47, they will run 24/7 on the server, and
- 0:50I can direct the routine tasks I want to
- 0:52assign straight to the team on the
- 0:54server. I also connected Cloud Code to
- 0:56control the entire system.
- 0:59So, even if I don't go into the server,
- 1:00I can control it as I wish from the
- 1:02terminal.
- 1:03Now, you know there is a tool called
- 1:05OpenRouter; you can use hundreds of
- 1:07models with a single key, and the
- 1:09platform even shows who is using which
- 1:11model and how much. For instance, in
- 1:13last week's list, seven out of the 12
- 1:16most used models were Chinese models.
- 1:18DeepSeek is in first place, of course,
- 1:20and I chose exactly five from this list
- 1:22. Five models are quite enough for me.
- 1:24By the way, I gave each of them a job.
- 1:25This Qwen 3.1 Pro is the most expensive
- 1:28and slightly better than the other five
- 1:30. That’s why I appointed it as the
- 1:32planner and the supervisor, in other
- 1:34words, the main kingdom. Then we have
- 1:36our GLM 5.3. It was built for coding, so
- 1:39we will have GLM do the main work. Now
- 1:41we have our DeepSeek V 4.1 Flash model.
- 1:44One of the most used and cheapest. For
- 1:46example, it gets assigned tasks that
- 1:48process a lot of text. This could be
- 1:49research, reading a PDF, or preparing
- 1:52specific cards.
- 1:53Now we also have MiniMax, you know, the
- 1:55M 3.1 Flash. I would like to state that
- 1:57this video definitely does not contain
- 1:58any sponsorship. Now, this Flash model
- 2:00acts as a second pair of eyes to check
- 2:02the cards. So, this is the one doing the
- 2:04tests, but who does the final audit? We
- 2:06also have Coder Next. This is also a
- 2:09code-focused and cheap model. It fixes
- 2:11their errors as well. Now we see five
- 2:13models and five separate companies; the
- 2:16writer and the supervisor won't be from
- 2:18the same company. By the way, there are
- 2:19also two new models that came out. We
- 2:21didn't choose them. I'm explaining the
- 2:22reason, too. One is Tencent's Hunyuan
- 2:24model and the other is Alibaba's Max
- 2:27model.
- 2:28Both of these only work on their own
- 2:30servers. So, any prompt I give or
- 2:32information I write goes directly to
- 2:34their company's server. This is actually
- 2:36not that big of a problem.
- 2:38For instance, the other five models we
- 2:39chose work perfectly fine on different
- 2:41companies 'servers as well. By the way,
- 2:43I had Fable research this. Fable also
- 2:45decided not to use them due to security
- 2:47reasons. But ultimately, since they were
- 2:49running within a server, they would have
- 2:50been working in a fairly secure area.
- 2:52Still, they made that kind of choice. I
- 2:54wanted to point that out. You can share
- 2:55your thoughts in the comments section.
- 2:57Now, my second decision is this. These
- 2:59models will not be running on my
- 3:00computer. They will run inside the
- 3:02server. Now, why did we choose a server
- 3:04?
- 3:04Well, I have my projects, files, and
- 3:06passwords on my computer, and I don't
- 3:07want to give a model I define full
- 3:09access to them.
- 3:10Moreover, there is no personal
- 3:11information inside the server. In the
- 3:13event of a potential bad package
- 3:14download or a virus entering the server
- 3:16, I will notice it and clear it
- 3:18immediately. My agents are in a separate
- 3:20box inside the server. The models
- 3:23themselves live in the cloud, and the
- 3:25agents on the server can call the models
- 3:27from there. The box logic is as follows
- 3:28.
- 3:29The program inside only works with the
- 3:31files in that box and cannot reach your
- 3:32files, but it can access the internet.
- 3:35It is open to the internet.
- 3:36Because it needs to call these models,
- 3:38we didn't put any restrictions on its
- 3:40internet access.
- 3:41Let's get to the part about how I
- 3:43connected them. Hermetic is running
- 3:45inside your server, and as you know,
- 3:46Hermetic is an agent program.
- 3:48You can open certain distinct profiles
- 3:50within Hermetic, and there is a separate
- 3:51agent in each profile; I opened five
- 3:53profiles here and connected a model to
- 3:55each one, and for starters, I just put
- 3:5610 dollars under that program. I went to
- 3:59the site, got a key, and set a spending
- 4:00limit. That's it. It's quite simple. We
- 4:03can manage five models with a single key
- 4:05. Now, there is a task board within
- 4:06Hermes, and you can divide the work into
- 4:08tasks from there and assign a profile to
- 4:10each task. And I actually write the
- 4:12commands, or prompts, from this screen.
- 4:15And inside this server, we placed
- 4:17certain packages that models can use,
- 4:19capability files, and also a separate
- 4:21explanation file so they recognize each
- 4:23other.
- 4:24Think of it like a rule file. And I did
- 4:26all of this with the local Cloud Code.
- 4:28I gave my server information to the
- 4:30local Cloud Code, and it set up the
- 4:31architecture inside. But of course,
- 4:32Cloud is not inside the team. It works
- 4:34independently within the server. Now let
- 4:36me give you a tour of the kingdom. By
- 4:37the way, I just made it. I created the
- 4:39new interface. Since it's new, the
- 4:41information on the right side might not
- 4:43have been updated yet. Now let's see how
- 4:45the Taht (Throne) thing works. For
- 4:46example, when I give a prompt like "do
- 4:49research," this prompt is forwarded to
- 4:51the correct agent. But who does all the
- 4:53checking, planning, and overseeing? It's
- 4:55Kimy. That's why Kimy usually provides
- 4:57guidance when we give it a task.
- 5:00Now, when I click on Kimy, I can access
- 5:02its rule file and terminal on the left
- 5:04side.
- 5:06Now, when I write anything in the chat,
- 5:08Kimy answers me automatically from here
- 5:10. The answer comes directly from that
- 5:12model. So, I can talk to any agent I
- 5:13want one by one. Now let's talk about
- 5:15what you can do with a system like this
- 5:16. First of all, this is a 24/7 operating
- 5:18system.
- 5:19It means it can work however it needs to
- 5:21when it connects to any external network
- 5:23from your Telegram, Discord, or WhatsApp
- 5:24.
- 5:26You're out and about, you want to
- 5:27perform a task, but you don't want to
- 5:29use Claude Code. In that case, for
- 5:30instance, I can write directly to this
- 5:31system. "Do some research for me."
- 5:33Competitor analysis, for example, is
- 5:34very good for this. Also, these are
- 5:36Chinese models. They are quite cheap. So
- 5:37, I know very well that they won't be
- 5:39able to run again once my limit within
- 5:41OpenRouter is reached. It also has
- 5:43another great feature. Let's say I
- 5:44didn't like Claude's performance. Then I
- 5:46can replace Claude with any model I want
- 5:48. Now let me explain a use case for this
- 5:50system, or how I use it myself. It works
- 5:52like this. I can provide any prompt I
- 5:54want here again. I can write a command
- 5:56in the chat section. I open one of these
- 5:57and say, "Do some research for me." For
- 6:00example, I say, "Go research the most
- 6:02viral posts on Twitter in the last 24
- 6:03hours for me." They don't know my
- 6:05account details on Twitter, for example
- 6:06. They don't know anything about me.
- 6:08Now, when we consider the architecture,
- 6:10even if I give the prompt, since Kimi is
- 6:12the CEO, the main agent, it
- 6:14automatically forwards the task to the
- 6:16agent that needs to fulfill it.
- 6:18Now I actually gave this prompt. Our
- 6:20prompt went to Kimi. Kimi split the job
- 6:22into four parts and then distributed
- 6:23them. It makes the decisions. You can
- 6:25think of it like a referee. DeepSeek
- 6:27went on the internet. It found 14 topics
- 6:29in one minute. Minimax opened these
- 6:31topics and their sources one by one to
- 6:32check them, and the total took four
- 6:34minutes. Then CLM wrote a report, and
- 6:36that took a minute and a half. Then Kimi
- 6:39audited the report and found five errors
- 6:41. For example, the dates of two sources
- 6:43were wrong, and Qwen corrected them
- 6:44afterward. Now I say open the result,
- 6:46and as you can see, the report is in
- 6:48front of me. 14 topics and 57 sources.
- 6:51By the way, I also checked the tests and
- 6:53results again with Claude Code to see if
- 6:54they did it correctly or not.
- 6:56It also confirmed that the sites were
- 6:58correct. In total, this process took 18
- 7:01minutes and cost 70 cents according to
- 7:03Hermes's calculation.
- 7:05So, I can say it's a great team effort
- 7:07for this research and daily routine
- 7:08tasks.
- 7:09I say it can use any MCP it wants. It
- 7:11can use any skill file it wants. This is
- 7:13very important because I don't download
- 7:15every GitHub repo to my local machine,
- 7:16for example. But these Chinese models of
- 7:18mine can download any GitHub repo they
- 7:20want. They can install any skill file
- 7:22they want. I will show you an example,
- 7:23for instance. There was a GitHub repo
- 7:25called Limo Opus Car. I had previously
- 7:27prepared a Reel for it on Instagram. It
- 7:29has great film styles in it, and your
- 7:31model knows how to produce video from it
- 7:33. I take the GitHub repo. I come to my
- 7:36local terminal here. I provide this
- 7:37GitHub link, especially noting that it's
- 7:39inside my Chinese models, and I say, "
- 7:41Install this for me.""Look at its
- 7:42content.""Using this, produce faceless
- 7:45content for me."
- 7:46I am looking at our Limo Opus Car
- 7:48results right away. I am monitoring the
- 7:50performance of all of them here. How
- 7:51much did they spend, for example, how
- 7:53much money did they spend setting up
- 7:54this GitHub repo? How many tokens did
- 7:56they consume while analyzing its content
- 7:58? What was the return on this? The team
- 8:00set up this repo inside the container
- 8:01themselves, and nothing was downloaded
- 8:03to my computer. It even produced a film
- 8:04, but the result isn't very good. You
- 8:06can see it for yourselves. It is already
- 8:08written on its page. This was originally
- 8:10designed for the Cloud. It states that
- 8:12other models might not produce good
- 8:13results. And that is exactly what
- 8:15happened.
- 8:15So, for setup and research, this team is
- 8:17very good at the job. It is also
- 8:18perfectly okay for scripts or perhaps
- 8:20different types of content production.
- 8:22But for this kind of film work or
- 8:24editing tasks, Cloud is still in the
- 8:26lead, as always. Now, let's have it do
- 8:28one with this rule. After recording the
- 8:31audio, I put it on Drive and gave the
- 8:33link to the team. Afterwards, I wrote
- 8:36the prompt as a command from here.
- 8:38Vertical, from this audio, in a
- 8:40powerless way. First, who distributed
- 8:42the work? The team worked for 25 minutes
- 8:45, and this is what they made. The
- 8:47subtitles appear at the same time as my
- 8:49voice. There is no video model. They did
- 8:51it all with just code, and it cost 1
- 8:53dollar and 10 cents. So, if you guide
- 8:55them correctly, you can get things done
- 8:57with these teams. Of course, the final
- 8:59review part can be done on your side,
- 9:01meaning locally. After all, the sharing
- 9:02will also be done locally. By the way, I
- 9:04think this kind of architecture can be
- 9:05improved. You can even add more models
- 9:08or use 2-3 different servers at once. I
- 9:10used a VPS as the server and utilized
- 9:13what was inside. By the way, this video
- 9:15is not sponsored. If you want, you can
- 9:17find the link in the description section
- 9:19.
- 9:19By the way, if there are other models
- 9:21you want me to try, or if there are
- 9:22parts where you think "you should have
- 9:23added this" to the architecture, you can
- 9:25mention them in the comments section.
- 9:26I will put the details regarding the
- 9:28video in the description. See you in the
- 9:29next video.
About this transcript
This page contains the full transcript of 5 Çin Modeliyle Krallık Kurdum: Sıfırdan Kurulum by Selma Kocabıyık, generated from the public captions YouTube serves with the video. The transcript has 2,150 words across 310 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.