Video15 min watch

NEW Qwen 2.5 Coder 32b BEST Open Source LLM?🤖 TESTED Beats GPT-4o Cursor AI Compose + Artifacts

In this video

  1. 00:00Introduction to Qwen 2.5 Coder 32B Model Overview
  2. 00:49Benchmarks and Performance Claims for Qwen 2.5
  3. 01:47Model Capabilities and Supported Programming Languages
  4. 02:28Running Qwen Locally with Olama and LLM Studio
  5. 03:00Using Qwen 2.5 with Cursor and OpenRouter Setup
  6. 03:38Open AI API Configuration for Cursor Integration
  7. 04:17Basic Code Tests with Qwen 2.5 in Cursor
  8. 04:39Python Fibonacci Sequence Test Pass
  9. 05:13Batch Script and SVG Generation Tests
  10. 05:44HTML/CSS Landing Page Generation Test
  11. 06:19Responsive Grid Layout Creation Test
  12. 06:54Medium-Difficulty LeetCode Challenge Test
  13. 07:15Hard-Difficulty LeetCode Challenge Attempt and Adjustments
  14. 07:48Creating a Task Manager App with ShadCN UI and Next.js
  15. 08:09Enhancing App Features and Dark Mode Testing
  16. 09:21Overall Impressions and Model Limitations Compared to Proprietary LLMs
  17. 09:53Closing Thoughts and Channel Community Invite
  18. 10:47Sponsored Segment: Stride Agents AI-Powered Appointment Setters

Links from the video

Transcript

2,202 words

What's going on guys Josh poock here now just over a month ago on this channel we covered quen 2.5 the quote unquote best open- source llm so we tested it before with the 1.5b and 7B sizes but now there's the 0.5b the 3B the 14b and the 32b coder model and just judging off these benchmarks this model looks insane comparative to models like GPT 40 or claw 3.5 Sonic they claim that it Rivals GP pt40 we're going to run some tests

And see what this model is made of and see if their claims and these benchmarks are actually what they say they are let's dive right into [Music] it all right guys so all links covered in this video will be linked down below the link to quen's tweet right here announcing that they have their quen 2.5 coder 32b instru model and here they claim that this model reaches top tier performance highly competitive or even surpassing proprietary models like GPT 40 in a

Series Benchmark evaluation including human eval mbpp Live code bench big code bench Mickey Val AER Etc as well as a link to their in-depth blog post going over these benchmarks that I would definitely recommend checking out if you want to deeper dive into this just to quickly go over some of this they say quen 2.5 coder 32b instruct has become the current SOA open source model matching the coding capabilities of GPT 40 while demonstrating strong and comprehensive coding capabilities it also

Possesses good General and mathematical skills they tested it in many different areas such as code generation code repair code reasoning so it supports multiple programming languages over 40 different languages scoring 65.9 on Mick eval all right it has a 128k token context window length uh under the Apache 2.0 license you can see all the different specs right here all right they tested it in different scenarios including cursor and artifacts and we'll showcase some of those in just a second there's also

A video right here that I would encourage you to watch and check out as well like I said this blog post will be down below so you can check it out more in depth now there's two main ways we can actually use this model one is the hugging face uh chat right here which I'll leave a link down below the other is the hugging face artifact which I'll leave link down below now if you have the right hardware specs you could

Also run this on uh locally using olama or llm studio all right so I'll leave this link down below as well you would simply just copy this command AMA run quen 2.5 coder 32b and you would just run that now you can also use this in something like cursor too so now to do that you would go to open router. a right here I'll leave a link down below and you can actually set this up within cursor now you'll see that

There also are some other quen 2.5 coder 32b instruct providers so we have deep infra we have hyperbolic and fireworks so you could also use some of these other providers as well I will leave them linked down below as well so hyperbolic here deep infra and fireworks but I'm going to be using open router just for Simplicity here so all you would simply do is just go into your cursor you could also do this pretty much anywhere where there is an

Open AI um base URL that you can change and you would simply just go to your settings and cursor you would go under open AI API key you would want to override the open AI base URL and you would paste this right here the open router. aapi slv1 you would save that and then simply just paste in your open router API key and then over here you would just actually go up here actually to the ad model and you would just

Simply copy the model name right here under with open router and paste that over into cursor like so as you can see I went down here and selected quen 2.5 coder and asked it if it is quen 2.5 coder and it said yes it is all right so we'll do some tests in cursor in just a second here now before I do those we're just going to go through some of these quick tests right here these are some tests that we've

Run on a number of different models on this channel so we'll run through through some of these basic ones some of them are a little bit more challenging than others and then we'll try some more advanced ones and try some incursor as well just to see the full scope of what this model is made of okay so I can say write a python function to generate the Fibonacci sequence to the 10th number okay so I'm going to run python f.p right

Here and boom we got the correct answer so that is a pass all right next we're going to say generate a bash script to create 10 files named a1. txt through A10 txt all right we created the file that generated now I'm going to run it and boom it created all the 10 files right here so that's a pass next is generate a SVG of a tree okay and it passed all right next I'm going to say generate me the pawn

Game using python all right so it's telling us to pip install P game right here I'm going to run that and while that's installing I'm going to apply this Pawn game right here into our file and save all right now I'm going to run python pong. py okay and we got a pong game right here not the best not the worst but it is a pass all right for this one we'll use the artifact on hugging face so create a landing

Page using CSS JS and HTML it should be a website of a beauty store that has a header Banner features and testimonials and checkout section make it look professional and very modern okay so we can see the artifact generating code right here we got our HTML and our CSS right here okay so we got our beauty store right now we can see features testimonials checkout so we got testimonials right here checkout page um features page page so definitely not the most

Visually appealing but it is a pass next we say create a responsive grid layout with HTML and CSS that displays four items per row on a large screen two per row on a medium screen and one per row on a small screen all right so we have four items here if we make it smaller there's two and then if we make it smaller there's one so that's a pass all right so it passed those basic tests with flying colors let's go

Ahead and try some more challenging tests right here right so we'll try out this medium difficulty one right here on leak code all right so we'll copy this right here and paste it right over here okay so it generated our code right here I'm going to go ahead and copy this and let's go ahead and paste it over here and we can run it okay it was accepted let's go ahead and submit this okay and it passed all right let's go

Ahead and try a difficulty level hard right here with a 49% acceptance rate design cancelable function okay so I'm going to copy this right here going to paste this right here and click Send okay so I'm going to copy the output right here and paste it right here let's go ahead and run it okay it accepted let's submit it oh okay so wrong answer okay I'll try one more time I'm going to copy this paste it back into here saying wrong

Answer all right so it did give us a revised implementation let's go ahead and try this one more time okay we're submitting it one more time here and it got the answer wrong again all right so next I'm going to say I want to create a very modern task manager app using shad and UI next J S15 app router using cursor okay so I'm going to run this Command right here to create the nextjs app okay so our next app is

Installed here let's go ahead and run this okay now I'm actually going to ask it to create this app using the composer right here and let's go ahead and click on submit okay so we got it set up it's not the definitely not the most efficient way if I was to actually set up an app um in terms of like the way I'm setting up this nextjs app but you know surprisingly it actually I mean not really surprisingly it did a

Pretty good job um to be honest I'll show you what it looks like and this is very very basic here as we can see but it is using Shad CM we got a simple task manager app we can go test right here add some tasks like so we can see learn examples okay so I guess this is from the original nextjs app it didn't remove that all right so I'm going to say add light mode or dark mode uh toggle to

My Shad CN app and make it way better with more features okay so it added this thing that shows how many tasks there are so if I add a third task it will go to three tasks and I can clear all tasks and it did add a button up here but the dark mode isn't working all right guys I think you get the point here all in all this model is pretty impressive in terms of Open Source models it aced some

Of those basic tests we did at the start and then it even did pretty good on those elak code tests you know it was struggling on some of the harder ones but it was good at the medium ones and in terms of using it in cursor with something like open router or one of the other um providers that I showed you it does pretty well in terms of generating code multifile editing Etc so all in all I would definitely say it

Isn't Claude Sonet 3.5 level or even GPT 40 level by any means you know it is maybe getting closer in terms of Open Source models but you know like I said for open source models it is definitely pretty impressive let me know what your thoughts are in the comments Down Below guys you know it does have some really really impressive benchmarks but after all we can't always just go off of benchmarks because so many models can just say they have the

Best benchmarks in the world and they're the best models but when you actually Implement them and test them and use them it may not actually back up those you know results so let me know what your thoughts are down below guys other than that if you're new to the channel we upload videos about AI Marketing sales business growth so if you like that type of content you got some value here like the video comment down below and subscribe to stay up

To date with the uploads also too guys if you haven't already joined our free community stri community.com I'll leave a link down below to our free Facebook group and Discord Channel and then also too guys if you run a business and you need help with systems Marketing sales operations or implementing AI agents into your business like AI appointment centers Ai call centers whatever the case may be book call down below at executiv st.com slapp other than that guys I will see

You in the next video keep hustling keep grinding and of course guys accelerate your stride take care are you tired of pouring thousands of dollars into appointment Setters only to watch leads slip away imagine having a team of elite sales agents booking qualified appointments for you around the clock no more wasted time on training no more frustration with performance and no more draining your budget on inconsistent and expensive call centers introducing stride agents AI powered appointment centers that work 24/7 never

Get tired and book appointments while you sleep trained on thousands of successful conversations our AI agents El performed human teams at just 1110th of the cost join the ranks of businesses that doubled their appointments and booking rates in just a matter of weeks don't get left behind in the AI Revolution visit stride agents.com now and transform your entire sales process with Cutting Edge AI technology it's time to accelerate your stride with AI agents

Latest from the blog

View all