I've been hoping something like this would be developed.. I'm so impressed.. Love the application of using it with different models simultaneously. And, of course, to run it all on my computer.. You are awesome.!
@@digitalchild I like knowing how it's working. Having control over the fine details is important to me and especially. The ability to use it with my small language model on my own computer. Is a game changer for me.
Holy shit I was almost having a panic attack by not being able to fix the error of OpenAI API key, and you did it very simple and easy! Thank you so much!
all great information and each of your presentations. It would really be handy if hardware requirements were mentioned for those of us who have mineral hardware so we know what we can dive into and what not. Thanks for all your work that's awesome!
You bet Moses - thank you! I typically don't mention hardware requirements since it all depends on the LLM you choose to use. If you want to use a 32b parameter model or bigger for example, you'd need something like a 3090 GPU. But then most computers would be able to run a 7b parameter model or smaller.
Please compare Swarn and n8n. Because one is code based and another is no code. So please analyse whicb one is better for us because as a no code user I want to understand what I am missing in n8n. Thanks for the video.
@Cole really really awesome content , How about some content on how can use all these individual component (be that n8n, swarm or ...) into a end to end use case. Understanding how it works and implementing these into a production ready env is completely different. I will be looking forward for content like these. Thanks again for the awesome videos.
You are so welcome! And thank you for the suggestion! A lot of my upcoming content is going to be focused on combining a lot of the tools I've been covering in the way you described.
Going back a few videos to ChatLLM, it has RouteLLM that interprets the prompt and routes it to the best LLM for the job. What does the group think about taking a broader view of this Swarm agent router implementation and building a chat request router in front of a defined group of Ollama hosted LLMs? Have it easily trainable by the user for their use cases and their questions. Forcing non technical users to learn the capabilities of each model through trial and error is, IMO, an unnecessary headwind to general adoption. PS - hat tip to the Ollama Team who had the forethought to provide OpenAI compat ability to tweak variables precluding Cole from having to fork.
I think you made a great video on the technical side next time you make a video can you start with what this does for the end user? For instance what I could even use this for.
I love this ability and appreciate you showing us how to do it. I am very curious on how someone could leverage this to create a secondary fund stream? I see tons of video's about code assist but nothing about how to utilize it to create funds.
You'll have to do like you did with Bolt, meaning, your forked version for the community with multiple LLMs, LOL. Like trying with Groq with LLMA 3.2 90b ect.. My idea is letting the main agent choosing the right and best LLM for simple and complex task with the tools if need ect. For example, taking a small model to scrape weather or stock price or adding tools with free API and return to a small llm for the result.
Haha yeah I might! Though a lot of providers like Groq don't support function calling with their OpenAI compatible API unfortunately. I love your thoughts here!
Haha it's tough to keep up, especially with the other Bolt.new fork content I am already planning! haha I'll have more updates on the project including this feature this weekend!
If I use two different models in a script and switch back and forth between them, will each model be removed from the video memory every time I switch, potentially leading to longer runtimes because both models can't fit into the memory at the same time?
Good question! I believe there will be latency because of that, which is definitely something to keep in mind. Typically when you are using multiple LLMs at once, one is much smaller and meant more for routing tasks while the large one handles the complicated stuff.
I did this myself. Called it llama swarm. I made it so mine connects to a SQL database to help keep track of progress. This way you can run with free model through open router. Helps shrink the token count. It writes files to the local computer as well.
I tried this out, but results were pretty iffy I tried to add on to this code by creating a "help" agent that would just read back some examples actions that could be taken given the table schema in the context. I figured that would be a reasonable kind of use-case in this scenario, given that a user wouldn't actually know what kind of data they can get out of a db. However, it was tripping up somewhat often, often times suggesting me to write a specific query rather than executing it on its own. Other times Just getting into hallucaination loops or saying I should google it. This is on the models you used here qwen2.5-coder 7b and qwen3b for the router
@ wow that’s awesome. It’s like Tony Stark in that scene installing different AI in his suits from disk. I expect future builds of windows will have integrated AI LLM in the OS.
My gripe with crewai is it is calling home. I wanted to work and test my local code without internet and couldn't because it couldn't call home, don't know if they addressed this or not but I work with sensitive data that I can't risk getting out so to speak.
I've been hoping something like this would be developed.. I'm so impressed.. Love the application of using it with different models simultaneously. And, of course, to run it all on my computer.. You are awesome.!
You have been able to do this for over a year with crewai and it’s far superior to swarm.
Thanks so much man!
@digitalchild True, but I love how simple Swarm is, part of why it makes such a great educational framework.
@@digitalchild I like knowing how it's working. Having control over the fine details is important to me and especially. The ability to use it with my small language model on my own computer. Is a game changer for me.
@@ColeMedin Yes, this is very true.
Holy shit I was almost having a panic attack by not being able to fix the error of OpenAI API key, and you did it very simple and easy!
Thank you so much!
You are so welcome!!
@@ColeMedin Thanx for all
Oh man, it's 0140 am over here... Now there's no chance of me getting any sleep tonight 😂 Thank you Cole for another excellent video😊
Haha wow well I hope you are able to sleep in a bit 😂
You bet!!
You are making better Engineers out of people, thanks for that.
I appreciate it!
all great information and each of your presentations. It would really be handy if hardware requirements were mentioned for those of us who have mineral hardware so we know what we can dive into and what not. Thanks for all your work that's awesome!
You bet Moses - thank you! I typically don't mention hardware requirements since it all depends on the LLM you choose to use. If you want to use a 32b parameter model or bigger for example, you'd need something like a 3090 GPU. But then most computers would be able to run a 7b parameter model or smaller.
Great, simple explanation! Thank you!
Thank you! You bet!
YES. This is fantastic.
Thank you! :D
Nice integration qwen2.5 is very smart for its size
Thanks and yes it sure is!
Great contribution 👍
Thank you very much! :D
Please create a frontend UI to run this project.... This makes this beautiful script look awesome .... Great job ....
I certainly might! Thank you!
Please compare Swarn and n8n. Because one is code based and another is no code. So please analyse whicb one is better for us because as a no code user I want to understand what I am missing in n8n. Thanks for the video.
I agree 👍
great question!
You are welcome! I love your suggestion and I've noted it down!
Hi real gourav, Can we connect?
@@resourceserp ok. what is it about?
Great video, you can also use groq as well as that also uses the OpenAI format
Thank you! I actually tried using Groq but it seems their OpenAI compatible API doesn't support function calling, so I was getting errors.
@Cole really really awesome content , How about some content on how can use all these individual component (be that n8n, swarm or ...) into a end to end use case. Understanding how it works and implementing these into a production ready env is completely different. I will be looking forward for content like these. Thanks again for the awesome videos.
You are so welcome! And thank you for the suggestion!
A lot of my upcoming content is going to be focused on combining a lot of the tools I've been covering in the way you described.
I recenly came to know about cofounder ai , i didnt saw it all but i think that is worth exploring and explaining to ppl
Yeah I've heard a lot about it recently, I agree!
Going back a few videos to ChatLLM, it has RouteLLM that interprets the prompt and routes it to the best LLM for the job. What does the group think about taking a broader view of this Swarm agent router implementation and building a chat request router in front of a defined group of Ollama hosted LLMs? Have it easily trainable by the user for their use cases and their questions. Forcing non technical users to learn the capabilities of each model through trial and error is, IMO, an unnecessary headwind to general adoption. PS - hat tip to the Ollama Team who had the forethought to provide OpenAI compat ability to tweak variables precluding Cole from having to fork.
I think you made a great video on the technical side next time you make a video can you start with what this does for the end user? For instance what I could even use this for.
Thanks for the suggestion! I totally agree that's important too
I love this ability and appreciate you showing us how to do it. I am very curious on how someone could leverage this to create a secondary fund stream? I see tons of video's about code assist but nothing about how to utilize it to create funds.
You bet! Could you clarify more what you are asking? Like how you could make a side hustle with this?
You'll have to do like you did with Bolt, meaning, your forked version for the community with multiple LLMs, LOL.
Like trying with Groq with LLMA 3.2 90b ect..
My idea is letting the main agent choosing the right and best LLM for simple and complex task with the tools if need ect.
For example, taking a small model to scrape weather or stock price or adding tools with free API and return to a small llm for the result.
Haha yeah I might! Though a lot of providers like Groq don't support function calling with their OpenAI compatible API unfortunately.
I love your thoughts here!
Surprised not to see a video after bolt.new video today 😅 Any update on installing a local project to your repo?
Haha it's tough to keep up, especially with the other Bolt.new fork content I am already planning! haha
I'll have more updates on the project including this feature this weekend!
@ColeMedin
Thank you so much for this video. Please let us know it possible to use any other LLM provider such as Gemini or Anthropic.
Yes you can do something similar to what I did in this video to use really any LLM provider!
Mr Cole, did you see that M4 Max MB Pro? ... by the way, the bolt folk project is awesome. I having so much fun with it
I'm glad you are!! What about M4 Max MB Pro?
need a video on Cofounder AI. its amazing.
So I've heard! I will be looking into it!
Nice vid! If you could do the same with Langflow or Flowise, which are more "production ready" than swarm, it would be really appreciated :)
Thanks and I appreciate the suggestions! I definitely want to do some content on both in the future.
@ColeMedin appreciate your vids Cole! I'm trying to find the best platform to do these agentic bots, so videos like these are surely helpful :)
If I use two different models in a script and switch back and forth between them, will each model be removed from the video memory every time I switch, potentially leading to longer runtimes because both models can't fit into the memory at the same time?
Good question! I believe there will be latency because of that, which is definitely something to keep in mind. Typically when you are using multiple LLMs at once, one is much smaller and meant more for routing tasks while the large one handles the complicated stuff.
Is it also possible to use this with Claude sonnet or Gemini?
Certainly!!
Can this be done in n8n? The SQL n8n agent I think it's not as good as the one in swarm...but maybe I'm wrong!
Yeah it certainly could be! Swarm honestly makes it even easier to set up though IMO.
Could swarm be added to the ai stack to work with n8n and webui?
It certainly could! That is one thing I am considering to add, though there are a couple great contenders for agent frameworks to add in!
I did this myself. Called it llama swarm. I made it so mine connects to a SQL database to help keep track of progress. This way you can run with free model through open router. Helps shrink the token count. It writes files to the local computer as well.
Do you have a channel / video ?
This is what I just looking for!!!
Escape from chat gpt api 😂😂😂
Haha I'm glad!!
I tried this out, but results were pretty iffy
I tried to add on to this code by creating a "help" agent that would just read back some examples actions that could be taken given the table schema in the context.
I figured that would be a reasonable kind of use-case in this scenario, given that a user wouldn't actually know what kind of data they can get out of a db.
However, it was tripping up somewhat often, often times suggesting me to write a specific query rather than executing it on its own. Other times Just getting into hallucaination loops or saying I should google it. This is on the models you used here qwen2.5-coder 7b and qwen3b for the router
Yeah I get it, smaller LLMs still have a ways to go as far as performance for these kind of things!
Do you know how can I use hugging face LLMs with agents? I couldn't find anything related to this on the internet.
I've actually had trouble using HuggingFace with function calling... much better luck with Ollama for me. I think their API is limited in that way.
Now there's Skyvern ...I can't keep up anymore lol
Are you downloading the AI models to your local computer?
I sure am! Using Ollama it's entirely local.
@ wow that’s awesome. It’s like Tony Stark in that scene installing different AI in his suits from disk. I expect future builds of windows will have integrated AI LLM in the OS.
You do not have to use Ollama, it is the exact way to use the LM Studio server as well.
True!
I’ve found crewai to be far better in terms of results and capabilities. It has always supported multi model agents including ollama.
Issue with crewai is that it requires LLM for every task it does... You cannot mix and match AI vs non AI tasks
My gripe with crewai is it is calling home. I wanted to work and test my local code without internet and couldn't because it couldn't call home, don't know if they addressed this or not but I work with sensitive data that I can't risk getting out so to speak.
yeah I know a lot of people who love CrewAI! I think Swarm is great for education though since it is so simple.
Auto groq was doing this months ago
True! I love how simple Swarm is though - it's definitely a superior educational framework IMO.
I played with a 3b model and my computer froze for 10 minutes 🙄
Yikes I'm surprised for a smaller model like that! What are your specs?
Zoom it dude
Wow! you are so awesome! I will deposit $100 USD to you..I just emailed you...
Thank you! I'll take a look soon! 😀