Workshop With Nebius: Using Open Source Models for Agents | AgentX-AgentBeats Competition

Who speaks in this episode?

• Sujee Maniyam (Main speaker)

What is discussed, and when

1. 0:00 – 5:00: Nebius, Sujee Maniyam, AI hackathon, Kimi
2. 5:00 – 10:00: Nebius, Token Factory, DeepSeek
3. 10:00 – 15:00: Token Factory, Nebius, Agnos, CrewAI
4. 15:00 – 20:00: Token Factory, CrewAI
5. 20:00 – 25:00: Nebius, Token Factory, CrewAI
6. 25:00 – 30:00: HTML, Nebius, AI hackathon, Discord
7. 30:00 – 35:00: Sujee Maniyam, Discord
8. 35:00 – 37:51: Sujee Maniyam, Discord

Notable quotes from "Workshop With Nebius: Using Open Source Models for Agents | AgentX-AgentBeats Competition"

• And today I'm going to talk about how to use open-source models with Nebius.

  about Sujee Maniyam · at 0:23

• So you can actually choose and host the data-- choose the data center that's geolocated and also for your data privacy as well.

  about Nebius · at 4:29

• So we have all the latest open-source models that we have seen, like Kimi, Lama, Qwen, DeepSeek.

  about Lama · at 5:58

• You can start with just changing two lines of code, and then we handle all the scales behind the scene.

  about Token Factory · at 6:46

• And we're also looking forward to see what you're going to build in this hackathon as well.

  about AI hackathon · at 8:43

• We have integrations with all the popular agent frameworks.

  about CrewAI · at 10:11

• And if you ever need support, you can come to our Discord server and ask questions.

  about Discord · at 10:58

• I'm going to say, implement a Snake game that's playable in the browser in a single HTML file.

  about HTML · at 27:18

Who and what is mentioned in this episode?

People

• Sujee Maniyam — Person · Main speaker · 7 mentions · first at 0:00
  Presenter leading a workshop-style talk on using open-source models via Nebius Token Factory, including demos, pricing guidance, and Q&A support for hackathon participants.

Companies & orgs

• Nebius — Company · Business focus · 13 mentions · first at 0:19
  Nebius is the company hosting the hackathon and presenting their AI cloud platform and Token Factory product.

Products & tech

• Token Factory — Product · Primary owned · 10 mentions · first at 5:15
  Token Factory is Nebius's enterprise-grade inference platform for open-source models, highlighted as a key product.
• Discord — Product · Support channel · 5 mentions · first at 10:58
  Discord is mentioned as a support channel for Nebius Token Factory users.
• CrewAI — Product · Incidental · 4 mentions · first at 10:11
  CrewAI is mentioned as one of the agent frameworks integrated with Nebius Token Factory.
• Lama — Product · Incidental · 3 mentions
  Lama is mentioned as one of the open-source models available on Nebius's platform.
• Qwen — Product · Incidental · 3 mentions
  Qwen is mentioned as one of the open-source models available on Nebius's platform.
  So we have all the latest open-source models that we have seen, like Kimi, Lama, Qwen, DeepSeek.
• Kimi — Product · Incidental · 1 mention · first at 3:08
  Kimi is mentioned as one of the open-source models available on Nebius's platform.
  So we have all the latest open-source models that we have seen, like Kimi, Lama, Qwen, DeepSeek.
• DeepSeek — Product · Incidental · 1 mention · first at 6:00
  DeepSeek is mentioned as one of the open-source models available on Nebius's platform.
  So we have all the latest open-source models that we have seen, like Kimi, Lama, Qwen, DeepSeek.
• Agnos — Product · Incidental · 1 mention · first at 10:11
  Agnos is mentioned as one of the agent frameworks integrated with Nebius Token Factory.
  We have integrations with all the popular agent frameworks.
• GPT-5 — Product · Incidental · 1 mention · first at 3:30
  GPT-5 is mentioned as a proprietary model benchmarked against open-source models.
  But the open-source models are very, very close behind.

Concepts & topics

• HTML — Concept · Framework · 3 mentions · first at 27:21
  HTML is mentioned as a markup language used in examples during the presentation.

Places & events

• AI hackathon — Event · Setting · 2 mentions · first at 1:01
  The AI hackathon is the event where the presentation is being given, focusing on using open-source models with Nebius.

Full transcript with timestamps

0:00 SUJEE MANIYAM: All right. Hello everyone. Welcome. Nice to have you all here. I hope you can see my screen and hear me OK. You can give me a thumbs up so I know you can. Cool. My name is Sujee Maniyam. I am a developer advocate with Nebius. And today I'm going to talk about how to use open-source models with Nebius. And a little bit about me. I'm in Bay Area, in San Jose, to be exact. And here are my contacts. You can contact me here if you have any questions, comments, or if you are local, you want to meet up for pickleball.

0:45 So any of them works. All right. I'll leave this for one more second. If you want to take a quick picture of this. So you can contact me later. All right. So let's dive right in. And I figured I'll start a little presentation for an AI hackathon with an icebreaker, some jokes created by LLMs. So I asked my LLMs, tell me a joke about agents. And so I asked two LLMs. And I'll put both of them up. And you can see which one's funnier. Or maybe they are equally bad. So I thought that was funny, just you know. Promotion one, I thought pretty good.

1:32 Anyhow, just a little icebreaker. Hope you all get a little fun out of that. And so let's talk a little bit about the open-source models. If you look at this chart right here, look at the timeline. This is going from 2022 to '24, which is last year. And this shows the closed-source and open-source models side by side. And you can see at the start, the closed-source models-- all right, the gap is pretty big. And as time goes on, they are both improving. You can see. But the open-source models are catching up pretty rapidly.

2:12 And towards the end of 2024, about a year ago, you can see they are very close to each other. But remember, this is only going to 2024, last year. And a year in AI is like an eternity. So this year, we have seen some pretty amazing models come up. And I just sort of like this one. So I just put this up here.

2:39 There we go. Yeah. So we have seen some pretty amazing models came up this year, especially this summer. And you can see some of these. And you can also see, these are pretty nice and pretty big models. And you can see there-- one second. Let me try it on. There you go. And you can see these are-- if you look at the model parameter sizes, these are hundreds of billions, 200, 400. Kimi is one trillion, so just crazy big models. And also, they are pretty good context window. So especially for agentic workloads, that is important.

3:18 And if you are wondering how they do, how well they perform compared to closed-source models, you can see. And here's a benchmark from artificial analysis. The top models are proprietary. You can see GPT-5, Gemini. But the closed-source model-- I mean, the open-source models are very, very close behind. And if you look at the scores, they are very, very close. So I always say. I mean, so maybe they're state-of-the-art, maybe it's still proprietary models. But the open-source models are very, very close behind. So what that means for us builders and developers is you can actually build applications, really sophisticated applications, using open-source models.

4:01 And that gives you a lot of freedom and flexibility and choice. And this kind of brings me-- I want to talk about Nebius very quickly. So you may have heard about Nebius. We've been pretty busy lately. We are a AI cloud. We are designed specifically for AI workloads. And by the way, here are our data centers in the US and Europe. So you can actually choose and host the data-- choose the data center that's geolocated and also for your data privacy as well. And one thing that stands out for us is, we actually control everything from the lowest level, from the hardware all the way to the software and orchestration layer.

4:48 What that allows us, it gives us pretty good control over the entire stack. And it gives you an advantage of price over performance, so just to give you a quick sense. So this is a quick look at the platform. You can see we build on GPUs. And then we actually build our own servers and own data centers. And we implement our own software all the way to the cloud layer. And the Token Factory, which I'm going to talk about in a second, is actually built on top of this. What that means is, it actually benefits from this AI cloud that is designed from the ground up for AI workloads.

5:31 So a Token Factory is basically what we like to call enterprise-grade inference using open-source models. So because we just saw how great open-source models are, and they are state-of-the-art, but how do you run them without having to worry about infrastructure? That's where we come in, Nebius Token Factory. And I'll show you a quick demo on this one. But I'll run through a quick slide. If you look at the Token Factory, basically, we have a full stack. So we have all the latest open-source models that we have seen, like Kimi, Lama, Qwen, DeepSeek.

6:03 We have all those hosted. You can also fine-tune the models. You can use the ones we have. Or you can adopt the more open-source models to your workloads. So we provide full fine-tuning capabilities. And also, we have things like tool calling, MCP, and also security compliance. These are all the things enterprises look for when they're choosing an inference platform. And yeah, what we call OpenAI compatible, so if you look at our API call, we use OpenAI APIs. And you can see here, I'm changing just literally two lines of code to point to Nebius endpoint.

6:40 So you can see these Nebius endpoints. And then we're specifying the API key. And that's it. You can start with just changing two lines of code, and then we handle all the scales behind the scene. And you can see we have pretty high rate limits. But what we say is, we have dynamic rate limits. We'll scale with the application. And if we need even more capacity, we have what we call dedicated endpoints that give you guaranteed capacity. Price performance. You can see, again, I took this from artificial analysis. So the green quadrant is where you want to be and give you good performance and good price.

7:15 And you can see, if you look at these benchmarks, you would see Nebius on the top, on the top tier. So what that means is, like, cloud is built specifically for AI workloads. So we give you pretty good price-performance ratio as well. So building agents with the open-source model, so now you've seen open models. I gave you a little bit about Nebius and Token Factory. So how do you build agents? And since most of you know agents, in this hackathon, I'll zip through this pretty quick. Agents are basically what we call autonomous, meaning they can operate without human intervention for the most part.

7:53 And they make smart decisions. They adjust, they tweak. They can adapt dynamically. So for us, if you want to build agents-- again, I was going to quickly, quickly run through this real quick. So they can actually learn from the environment, think for themselves, and adapt their workflow. So this is different from a regular workflow. For example, if you had a regular workflow defined, that is pretty static. I mean, you may have seen regular workflows designed by people. You'll say, OK, if this, do this, if not, do this.

8:25 Those are already predetermined. But an agent, on the other hand, can adjust the workflow dynamically. And that's what makes them special. Yeah. And again, all the use cases-- customer service, health care, finance-- we started to see more and more agents deployed in the real world. And we're also looking forward to see what you're going to build in this hackathon as well. Another thing I'd like to talk about is, which models would you choose for agents. And again, there's not, like, a one model that works for everybody.

8:59 But there are a few things I would recommend you guys look for. One is, roam agents are all about tool calling. So how is your model with tool calling capability? Does it do well? And we usually probably want a large context window. For example, especially coding agents or analyzing large document bases, you want to have good context window. That's important. And also, usually, the reasoning models tend to give you reasoning for the agents. Because remember, agents can think for themselves, so you want a good reasoning model to facilitate that.

9:31 Again, you can look at the Berkeley function calling benchmark. They update the benchmark pretty regularly. You can see what are the top models and then start with those. And here are a few I picked out from the benchmark. You can start with these. And remember, we have all these open-source models hosted. So you can switch around and try different models and then see which one gives you a good performance. So you're not locked into anything. And that's the beauty of open-source because you do have a lot of choices.

10:04 And when you're building an agent, you would probably end up using an agent framework. And we have covered there. We have integrations with all the popular agent frameworks. You can see Agno, CrewAI You probably recognize these names. You're probably already using a few of these. And so a Token Factory actually integrates with all of these. And I will show you some examples on this one in a second. So here's a page you should probably take a picture or a screenshot of. And so this has all the information you need to get started.

10:38 By the way, if you're part of the hackathon, we actually sent you a Nebius credit code that you can use. I'll show you how to use it in a second. These are the URLs, So you want to go to tokenfactory.nebius.com. And then we have a lot of sample code for you in our cookbook. And if you ever need support, you can come to our Discord server and ask questions. Yeah. So hopefully, everyone had a chance to take a picture or a screenshot of the screen.

11:13 All right. So let me show you some code to get you guys started. First up is, how do you apply the promo code. By the way, the promo code is actually sent out by email. So check your email, maybe check your spam folder. You should see the promo code from the hackathon organizers. I'm going to show you how to apply the promo code real quick. Let me know if you can still see my screen.

11:49 OK cool. So once you log into Nebius Token Factory-- let me make this a little larger. All right. Just log your login. It's very easy to log in using Google or GitHub. And then once you log in will see a pop-up button right here where I'm pointing.

12:15 And if you have the code with you, you can do this right now. Just click on the top button and then click with the promo code. Then you enter the promo code. That's it. And then as soon as it activates, you will see it here. The promo codes given to you guys only have one activation, so please don't share them. So the code you have is unique. You can see it's a unique code. It has one activation. So we emailed the code one per person, so careful sharing this because only one. So as soon as one person activates, that's it.

12:54 The code is used. So I want to make sure. Is everybody able to activate? So any questions, comments? Let me look at the Zoom real quick. OK good. So let me show you a few things for you to get started with the Token Factory. So here are the models. So let's start with the models here. You can look for models, like text-to-text model. This is the model that most of us end up using. You can filter those. And you can see, we have gpt-oss, Qwen. All these state-of-the-art open-source models are here. We also have embedding models if you want, text to image.

13:46 Text to image, meaning you specify text, and it'll create an image for you. So all that. And also, you can quickly filter the models. For example, let's say, you are looking for Qwen, you will see, just type here. And then we will filter this for you. Yeah. And for each model, you will see what we call a model card. You'll see the model, how much it costs. So hopefully, you don't have to worry about this because at this point, you should have credit activated. And also, you can see where it is hosted and what are the capabilities.

14:23 It's good for code, math, and can do JSON, things like that. Very cool. So very easy to navigate. And one way you can try these models is, we have this Playground. It's very easy. For example, let me look at gpt-oss. Let me go to the Playground. Click on Playground here. And then just very simple, just punch in your prompt. So let me actually copy-paste this here. I'll just say, try the Python code to convert from CSV to JSON file. And that's it. That's my prompt. And you can see how fast it is. So that's my Python code right here.

15:15 And it's done. So yeah, so you can play around. You can play around with any of the models, experiment. You could even test function calls. Like for example, if you're trying to do tool calling, and if you have custom function, you want to experiment, you can click on the functions, and click on Add function and follow the prompt and add functions and test to see if your model can access the function. Super easy. So let me pause you for a second. Any questions or comments? So all the models, you can go to the Playground.

15:47 Oh, by the way, so that's a Playground. If you want to know the code, you can simply say, View Code, and there we go. Just copy-paste. And you can use Python, JavaScript. And then you can be up and running, so super easy to use. And again, like I mentioned, there's only two settings we had to change, from an OpenAI client, you point to a base URL to token factory and also API key. And how do you get API keys? Very simple. Here you can click on Get API Key. And then you can give it a name. And just like other API keys, you give it a name.

16:30 I recommend giving it a descriptive name. So for example, you want one for hackathons or create agents hackathon. And then just say, Create. And then you want to copy that. By the way, you should copy the key and save it. Because remember, once you close the dialog window, you cannot get back to the key. And it's a security precaution pretty much all of the vendors do. So once you create a key, just make sure you copy it to your machine. That's it. And that's your API key. So if you look at the API key, that's you're talking about.

17:06 All right. And also, once you're done with this-- again, so let me just put this one code when you run the prompt That's a system prompt. And I want to modify that. And then after it's done, you can see the token speed and how much time it took the whole thing. So at the end, you'll see-- there we go-- how much time it took, what's the token speed-- so this is pretty good, actually-- and all that. All right. So that's how you use a model in the console and also using the API. Again, just click View Code. And that's it.

17:54 And then we have plenty of examples in the cookbook. So again, the URL here, you guys saw the URL here. So it's called token factory cookbook. If you go here, and you will see we have lots of sample code here for you. And whether you want to use our models, APIs, you want to use a RAG, probably, agents are probably of interest. So you can click on the agents here. And you can see, we have plenty of agent examples. And another thing, and then it's easier to navigate here. So you can start with a simple agents. Also, you can start by framework.

18:35 So for example, let's say, you're using Agno or CrewAI. And you're like, hey, show me a sample code for CrewAI. And you can just navigate very easily. So let me say, the research agent is what you want. Click on this. And you can just run this. Actually let me show you real quick.

18:57 So quick, quick check. Can you see my screen OK? I'm in the terminal. OK. Sure. Yep. All right. Thank you. So here's my CrewAI research agent. So it's going to just do a very basic research. Let me show you the code. Again, so this is my CrewAI research agent. And you can look at the code here. So you can see I'm importing from CrewAI. And again, so I'm using this model, Qwen3-30B. And you can see I'm punching in the API key here. Yeah. And basically, it's a very simple research. So I'm saying, you are a researcher.

19:50 You're going to do some research about groundbreaking technology in AI, so no big deal. And then how you run this? Very simple. So in all our code, we actually use uv for our example. So you have to do is just go to the directory and do uv sync to basically install the dependencies. Oops. Sorry. I am on a different-- my bad. So right here. All right. So all you do is like so, I'm the right agent, you make sure. Yep. And you'll be saying, Done. And then all you have to do is just do a uv and pick up instructions here.

20:36 This does usually run Python. That's it. So let's run this real quick. And now hopefully, you should actually do some research about our AI trends and come back with an answer. Let's see. And we are using the CrewAI agent library. There you go, up and running.

21:22 And also, if you have any questions, please put go ahead and type in your questions. And we will go through them towards the end. Yeah. And while it's doing this, let me show you a couple of other things.

21:40 There you go. It's done. So what did I come back with? So next big step is embodied AI system with continuous lifelong learning. So yeah, it came up with some research and came up with some topics. And also, you can see the usage how much it used up, how many tokens it used, and all that. Anyway, so you see how easy it is. And our examples are super easy to use and run. So what you can do is, you can just clone the repo or fork on your own and then just start writing some code, start adapting our code for your needs.

22:21 Also, agent demo, career research. Another thing you might be interested is, I use some open-source models as part of my development also, not just using it as an agent. So I actually use them quite a bit for writing code. So if you can see my VS code, you can see I actually have client installed. So client is an open-source coding agent. And they actually support Nebius models. So you can see here. So I installed client. Again, you can find them here in the marketplace, C-L-I-N-E, Cline. And then if you look at here, see how I have Nebius model set up.

23:04 So I choose a provider as a Nebius studio on my API key. And I choose then encoder as my model. And this is a pretty capable model especially for coding tasks. And you can see it's called Coder. So it's specifically designed for coding tasks. And let's give it a try. So I am actually in this. I have this project, OpenAI Studio project open. So I'm going to go and say, on a very simple term, I'm going to say, summarize this project for me. Right So what you do is, you should basically actually go to Nebius. You can see the progress here.

23:47 It's kind of reading some files, trying to figure out what this project is all about. And they do a very nice job on the top. You can see the API calls and how much tokens you are using. And this model, you can see that's a pretty large token window, 260k token window. And then so it can accommodate large files. Anyway, so I think there's a summary. There you go. You can see project overview. There's a Nebius Token Factory. It's a collection of guides and examples, yada, yada, yada. And then it gives me the breakdown of how the codes are organized agents.

24:27 You can see all the examples, RAG and all that. So pretty cool. So you can see I'm using the open-source model posted on token factory with line running VS Code. And you can see how much progress it made. We didn't even crack, just only 10% And you can imagine. So you can actually use this with large code bases. So yeah. So I love to see how you guys end up using these models in your day-to-day workflow. Let's see what else I want to show. Oh yeah. And another thing we actually do is, we actually sponsor and host hackathons.

25:11 And as part of our community effort, you can look at our community page. You can see contributions people are making. So these are contributions, apps and tutorials people are building. So if you have some cool app that you want us to feature here, please reach out to me. And you can see some cool stuff here, help desks. And also, we actually have events. And these are some of the events we hosted. So we actually host hackathons. And then we sponsor hackathons. We also list the Nebius prize winners and also other noteworthy projects.

25:49 So hopefully, I'm hoping, I love to see some pretty cool projects out of this hackathon. And we will list them here and give you a good exposure. Yeah. So check it out. So let's see how we're doing on time. All right. Very good. Oh yeah. So one more thing I want to mention. So when you sign up, I think when you sign up, you get a dollar, I think. But for agentic workloads, definitely activate using the code we have given you because that'll give you millions of tokens, trust me, more than plenty for you guys to do your work.

26:31 And also, we are on Discord. Let me put the slide up. We're on Discord here. I'm on the hackathon Discord. And also, we have one Discord. If you have any questions, you can post. In either one, you can tag me so I get notified and happy to help you guys out. All right. So let's see. Oh yeah. So let's end with the one fun demo. And what I have done is-- so you guys have seen I can actually use the console. So let me actually use gpt-oss and say, Go To Playground. And I'm going to ask you to say-- this is something I just do just a little demo.

27:18 I'm going to say, implement a Snake game that's playable in the browser in a single HTML file. Why not? And another cool thing you can do-- so here I'm using gpt-oss. I'm actually going to compare with another model. So you just saw what I did. You can just click on the Models for the Playground. And click on the Compare button right here. And you can compare two models side by side. And let me say, I select Coder, Qwen3 coder. And I'm going to give them the exact same task. I'm going to say, write me a single HTML file game of the game Snake, very simple game.

28:05 They should be able to do it. And here is a prompt. So let me see side by side what happens. And you can see they start writing the code. So anyway, so this way, I can see them side by side how they do. I will let them run. I actually did this earlier and then saved the output. So you can-- let me do a quick demo here. And this is gpt-oss. You can see from my screen name. All right. And this is the game right here. It's not too bad. So I can start playing. You can see here it's a snake game. So I see whatever red blob you tend to grow.

28:48 Anyway, you get the idea. So this is gpt-oss. Not bad. Single HTML file, you can see, single HTML file right here. So how did the Qwen3 Coder do? So let me bring it up. Refresh this. So this looks a little bit more substantial. And I just put the headings on my own. I just did the heading, just so I know which one's which. And you can see here, let's start playing. Let me see here. Anyway, so obviously, I'm not very good at this, you can see. But it's just a game. But it's pretty amazing. Just with a single prompt, you can see they can create these games in a single prompt.

29:35 How awesome is that? Anyway, so hope you enjoyed the demo. It's like a little fun demo I use. When you're trying a new model, say, OK, now, can you create a snake game with one prompt. There you go. To summarize-- let me put this up one more time. So there's a slide. And take a picture. Take a screenshot. Everything you guys need to get started is here, tokenfactory.nebius.com. And we have plenty of examples to get you started. And also, here are the code to get going. And if you run into any issues, any feedback, any suggestions, you can find us on Discord.

30:14 We love to support you. And also, remember, like I said, hopefully, I love to have you list your projects in our community page. I'm really excited to see what you're going to be building. All right. That's it from me. Thank you so much. I'll take a break to see if there are any questions, any comments. Go ahead, please. SPEAKER: Thank you so much, Sujee. I think we do have a few questions. SUJEE MANIYAM: Yes. Let's go ahead, please. Yeah. SPEAKER: Yeah. So one of them is more of a logistical question. The question is, when I go to Top Up, it asks for a country and account details.

30:52 My country isn't included. Is it OK if choose another? The country is Colombia. SUJEE MANIYAM: Right. So yeah, good question. Because since I had my account already set up, so you didn't see me. Yeah. So the very first time you clicked Top Up, you have some basic details because now, billing is involved. So yeah, just fill it out the best you can. Just make sure, you don't need to enter a card. You just want to select the promo code and then punch in the promo code. Because sometimes people see the card, and they're like, oh, no, no, I don't want to enter my credit card number.

31:25 No, we don't need the credit card, not just yet. So you can use the promo code and then do it. But yeah, we do need some basic details from you, like the country and things like this. Yeah, do the best you can. Yeah. Thank you. SPEAKER: And the next question is, let's see, can you connect external knowledge sources into the playground? SUJEE MANIYAM: Oh, interesting. So OK. Answer is yes or no. So in the function call, remember you can have a function defined. And this function can be some function. You can pretty much invoke any other function.

32:07 Give it a try and let us know. Typically, the way I have done this is by exposing a tool. For example, if you have a tool, just export it as an MCP. And then your model can connect to it. That's probably the easiest way you would want to do it. But this function call can give you some basic functionality to try it out and test your basic functionality. So yeah, so give it a try, please. And let us know because we do want to hear your feedback. But eventually, you actually want to migrate to the code. And once you have code, then it's pretty easy.

32:38 Your code, you can export using MCP or whatever. And then you can hook into any MCP call with-- that's probably the path you will end up choosing. But do let me know. Yeah, I'd love to know how you are planning on building this. SPEAKER: All right. Perfect. And I think we just have one more question right now. If I'm trying to minimize cost during development testing, how do you recommend designing a small enough test while trying to implement the benchmarks? Some of these benchmarks use a lot of tokens per eval. SUJEE MANIYAM: Great question, actually.

33:11 So here's the thing. And I'll show you the most obvious one. If you go to our model page, you see the model card. In the card, we are pretty upfront about pricing. And so again, look at the gpt-oss pricing, like input and output token right here, like, $0.15 per million, $0.60 per million output tokens. Remember, typically, the output tokens end up costing you more. Why? Because the model has to work, think, write, and produce the token. So tokens tend to be more expensive. I mean, output tokens tend to be more expensive.

33:45 Look at the Kimi K2. You can do the comparison. You can see it's almost four times more expensive. Yeah. So what I would recommend is, you don't need to go to the top-of-the-line model when you're kind of experimenting or you're just running like a bunch of tests. You can choose, maybe, a lighter model, a cheaper model. Go with that. Build out your harness, build out your tests. Do a few test runs to make sure it's good. And you can see the pricing. A lot of the models here, some of them are expensive, some of them are not.

34:16 So choose one that's reasonable, gives you reasonable performance. And then once you're doing the final testing at the final stages, then you can switch to a really high-end model that'll cost you more. But yeah. But then the credit we are giving you, I think it's, like, $50. And that should give you pretty good amount of tokens for you to work with. Because remember, this is the price per million tokens. So $50 will give you 50 million tokens. That's pretty good. But yeah. So do let us know what your experience is like.

34:47 This is one option. And also, another thing if you're using coding agents, another thing you can do is-- like, most of them will have this Plan versus Act mode. So Plan mode is just information gathering. So that will probably end up using a little less token space. So you can switch between these two, and then see, kind of play around a little bit there. Yeah, so there are some pointers. But yeah, also post on Discord if you have any questions. We'll give you some suggestions. Good question. Yeah. SPEAKER: OK. And I think one more question is, can I import my own data set to fine-tune a model?

35:25 SUJEE MANIYAM: Very good. Yes. I am glad you asked me this question. Let me actually show you. Can you still see my screen? Yep. Yeah. Perfect. So you want to go to Fine-Tuning. And then fine-tuning is super easy on the platform. So again, Fine-Tuning, so let me just walk you through real quickly. And here's how you upload your data, simply drag and drop. I can-- up to 5 gig, I support it. I just upload the data set. That's step one. Step two, you choose the model that you want to fine-tune. You can pick the models here that we support.

36:06 Now, some of them, almost you can do one-click deployment. So you can start with that. But you can see there are a lot of models supported here. Just choose a model you want to go with. And then you can tweak, fine-tuning parameters. You can choose lower or full fine-tuning. And then also, you see the pricing. You see, hey, if you want to fine-tune, tuning a million token, this is what you're paying. And then next, and then basically, just honestly just simply say, Create Job. And that's it. So the fine-tuning job will be created in the background.

36:40 It'll run a few minutes or in a couple of hours. And then at the end, you will be notified. And you have a new model. And your models will actually show up under Private Models. So then you can just deploy. Yeah. So yeah, do give it a try. And if you run into any issues, let me know. We'll help you along.

37:08 Very good. Any other questions? SPEAKER: I think we're winding down on questions. SUJEE MANIYAM: OK, good. All right, guys. I will put this one up one more time, just so you know. So if you missed it, I want to make sure everybody has this. But remember, we are on Discord. If you have any questions, you can tag me on the Discord channel and happy to help you. And also, let me put up my contact real quick. So yeah, so you can connect with me here-- email Twitter whatever. if you need to, say, connect with me or send me a message, happy to connect with you.

37:43 And then, yeah, I'm very excited to see what you're building. And good luck to all of us. Thank you.

Questions about this episode

Who speaks in "Workshop With Nebius: Using Open Source Models for Agents | AgentX-AgentBeats Competition"?

Sujee Maniyam (Main speaker).

What is discussed in "Workshop With Nebius: Using Open Source Models for Agents | AgentX-AgentBeats Competition"?

In order: 0:00 Nebius, Sujee Maniyam, AI hackathon, Kimi; 5:00 Nebius, Token Factory, DeepSeek; 10:00 Token Factory, Nebius, Agnos, CrewAI; 15:00 Token Factory, CrewAI; 20:00 Nebius, Token Factory, CrewAI; 25:00 HTML, Nebius, AI hackathon, Discord; 30:00 Sujee Maniyam, Discord; 35:00 Sujee Maniyam, Discord.

What is mentioned in this Berkeley RDI Center on Decentralization & AI episode?

13 people, works and ideas are mentioned, including Nebius, Sujee Maniyam, Token Factory, AI hackathon, Discord, CrewAI. Each one on this page links to its own page with every episode that mentions it.

How long is "Workshop With Nebius: Using Open Source Models for Agents | AgentX-AgentBeats Competition"?

The episode runs 38 min; the transcript is about 5,845 words in 60 timestamped sections.

More from Berkeley RDI Center on Decentralization & AI

• All analyzed episodes
• People mentioned on Berkeley RDI Center on Decentralization & AI
• Topics mentioned on Berkeley RDI Center on Decentralization & AI