Transcription
Alibaba just released Quen 3.6. And the part that got my attention is not just a model. They also released something called Quen code, which is an agentic coding harness. Think Claude code, but open source. It reads your files, edits code, runs terminal commands, and iterates on its own work.
Now, I use Claude Code every day. I pay $200 a month for the Max 20x plan. And I'm not going to sit here and tell you that this will replace Claude Code. I know many YouTubers like saying this and making big promises, but Claude Code runs frontier models, the best models out there, and it cost real money for a reason.
But I'm a big believer in open source, and I keep wanting local models to get good enough for real work. So, every time something new comes out, I test it. But in this video, I'm going to install it from scratch with Olama, wire up Quen Code, and then throw three real tasks at it. A 3D aquarium, a designer portfolio website, and a full Mac OS style desktop. Let's see how close local has really gotten.
I put the full setup guide and all the commands on GitHub for this video. Link is in the description so you can follow along or set it up yourself later. Okay, let's set it up.
First step we need to do is we need to install Olama. If you already have it, you can skip ahead. If not, go to the Olama website. You can click on download on the top right or use this command to install Olama. I already have it installed, so I'm going to skip this one.
The next one is you need to go to models and find the right model, Quen 3.6. And the way you can pull this model is run `ollama pull llama3.6`. So copy this and if you haven't downloaded this already when you run this command it will start pulling the model but because I already downloaded this model earlier I can just start saying hi and you can see how fast the model is. Now I'm on a 128 gigs RAM M3 Mac with Apple silicon but this is really fast.
Now here's where it gets interesting. We're going to connect this model to an agent, not just a chat window like Olama. An actual agent that can read your files, write code, run commands, and iterate like Claude Code does. The tool is called Quen Code. It's open source from the same Quen team. So, I'll go to the Quen Code GitHub and I will copy the command to install the Quen code. Once again, it's like Claude code but open source. and I'll paste this command.
Okay, Quen code was already installed on my machine. So, it's going to start running Quen code. Now, we need to tell Quen code to use our local Olama model instead of calling a cloud API or the cloud coding plan. So, we don't want to use these. We want to use a local model. The config lives at `~/.config/quencode/quencode_settings.json`. And the setting looks like this. This is going to be on the GitHub repo that I'll link. But I'm going to copy this settings.json content and we're going to create a file and paste this.
The key things here: base URL points to the localhost:11434 which is Olama. The model ID matches exactly what we pulled. The API key is just the word "ollama" because Olama doesn't check keys. But the SDK requires something. And I'm setting the context window to 64K. And this context window thing is actually the number one gotcha I see people hit. Olama's default is only 4096 tokens. Your agent will look like it's forgetting things mid-task and you'll think the model is broken. It's not broken. It's just truncating your context silently. So let's also make sure we're changing the context length on Olama settings when we're serving Olama.
So I'm going to save this file and then I'm going to serve Olama and the moment of truth. Let's launch it. Okay. Now I'm going to paste this command. It says `quencode --model "openai/llama3.6"`. Now it doesn't mean that you're using OpenAI. It means Quen code should be using the OpenAI compatible SDK to talk to your local model. Olama exposes an API endpoint at `localhost:11434/v1` that speaks the same protocol as OpenAI's API. So typing "openai" just tells Quen code use the OpenAI SDK format and the base URL in your settings.json redirects it to Olama instead of OpenAI servers. It's the protocol, not the provider.
All right. So once you do this, you'll see that Quen is open with Quen 3.6. It says API key, but once again, this is just your local model. Now, I can ask a question like, what is 2 + 2? Okay, this is a simple math question. 2 + 2 equals 4. That's a good start. That's cool.
Now, let's actually test the agent loop. This is what separates a chat from an agent. Can it read files, write files, and run commands? So, I'll say "create a file called hello.txt with the content 'Quen code plus works', and then read it back to me." Okay, I'm going to say "always allow." The file has been created. And yes, indeed, hello.txt exists. It wrote the file. It read it back to me. Agent loop is alive. And this was pretty quick. So, this is the free Claude code alternative running on my laptop. But can it actually build things? Let's push it a little bit.
I'm going to give it three tasks, each one harder than the last. Let's see where it breaks. The demo number one is let's ask it to build a 3D aquarium. So, I'm going to give this prompt to it: "Build me a 3D aquarium as a single HTML file using three.js from CDN. Glass tank with water. At least five fish swimming around with simple animation. Some plants or coral at the bottom. Soft lighting. Make it interactive so I can rotate the camera with my mouse." So, let's see how Quen does on this one.
And it's been about 1 and a half minutes and it's doing its thing. And while this is working, you know, just to show like comparison between this and Claude code, why don't I just copy this and I'll open Claude and I'll paste the same uh command and I'll say name it uh fish_plot.html. I want to make sure this is separate from um from Quen. So once again, this is Claude code. We already have Quen at work. It's been 2 minutes and 36 seconds. Okay, so Claude, it's still doing its thing. Just want to show you like how long Claude takes, right? Keep in mind this is using cloud models in big uh GPUs, so you know it's obviously going to be fast. And also note that, you know, I just shut off a bunch of uh stuff on my local machine. So maybe that was slowing down quite a bit, right? So I want to be super transparent that I'm running this on my local machine with a bunch of other apps which I just closed. So okay, so just so you know, Claude Code is also taking a while to do this.
Looks like Claude is done. So Claude was done in 1 minute and 28 seconds. Okay. And this is what Claude built. Uh it is like an aquarium. I can see fishes. I can move and I can zoom. Pretty good for one-shot uh implementation of an aquarium. Now I'm hoping Quen can do the same thing. It's still working. All right. Looks like it has come up with the full HTML. It's 494 lines. I already see it. Oh, wow. So this is Quen. So, this was Claude and Quen did, you know, really good job. In fact, the fishes, I would argue, are better. Um, and also like it's it's way more clear. The the Quen version is really clear. I can also zoom exactly like I asked. So, you know, I'm super impressed with this. You know, it just once again, this is a local model on my own machine. And look at what Quen did. Obviously, it was a bit slower than Claude code, but my laptop was running a few other things and it's also my local model, right? It's not a big model in the cloud that is served in the cloud by Claude or Anthropic, right? This is a local model and fish_aquarium is from Quen. This is very impressive, guys.
So, I'm going to keep on moving. Let's say this: "Build me a portfolio website for UI designer as a single HTML file. Da da da da da." I'm also going to give this to Claude. Name it `claude_portfolio.html`. Right. I want to make it separate from what Quen does. All right. So Quen is already working. 24 seconds. "The user wants a portfolio website for UI designer as a single HTML file. Da da da da da."
Okay, so Claude took 2 minutes for the website to complete and this is what Claude came up with. Looks pretty neat, you know. It looks pretty neat. Yes, this looks pretty neat. Has this thing going on. Now let's see what Claude, Quen does. Now one thing to be clear is I've installed a bunch of skills and plugins on Claude code which I haven't done on Quen. So you have to keep that in mind as we evaluate these uh outputs. But uh Quen is still working. 3 minutes.
So after about 10 minutes, Quen 3.6 built the portfolio website and now let's look at how it looks like. And the Quen website looks like this. You know, it looks pretty good. Looks looks looks pretty good. Now, if we compare that to Claude, you know, I I don't see much of a difference. You know, maybe you see like some differences. Um, obviously I I like, you know, the the big font in Claude's design, but everything else, you know, feels kind of similar to me, right? you know, it has this revolving text thing. Um, obviously once again, you know, Quen is a lot more smaller model uh running locally. Uh, so you know, Claude might win this one just by a little bit, but I'm very impressed with uh what Quen can do.
Okay, one more test. I'm going to start Quen again for a new session. And this time we'll we'll go big, right? We'll we'll try to go really big and say something like "Build me a Mac OS style desktop environment as a single HTML file. Top menu bar with clock. Da da da da da you know." So we we'll we'll give this prompt to um Quen. Now this should take a few minutes. At the same time we'll give this prompt to Claude, so that we can compare the outputs one more time.
Okay, so Claude, let's see how Claude does. All right, Claude came back with the Mac OS. It has a notes app, a calculator app, a files app which I can click and it gives me um these these pictures and stuff. But once again, this is the web app simulating a Mac OS. So this is drawn by Claude. Pretty quick, like 2 minutes and 27 seconds. It took that long for Claude. Now let's check in on Quen. We know it's going to be slower. It's local LLM, but let's see how Quen does.
Okay, after about 10 minutes, minute minute-ish, it's now asking for write permissions. Looks like it created a 1,200 line, almost 1,300 lines of code. Let's see how it does. So the file is created. It's called `mac_os_desktop.html`. All right. And this is Quen's work, by the way. This was Claude Code. And look at this. So Claude Code just gave me like icons with, you know, they look they all look the same, right? Quen gives me actual like gearbox music. So I'm already impressed. Let's look. Let's look at the files. Um, interesting. So, looks like the the icons itself, they're not clickable, but it does look like Mac OS. Um, so, you know, it it it made it look nice on first glance, but the icons themselves are not clickable like Claude's output. So, uh, the first feel is really good. Maybe we can get it to be better with some additional prompting. I'm sure Quen can do it, but on first feel, you know, it it looks good. It is getting there, but it's definitely not working as expected.
All right, so here's my honest take after using this for a day. A year ago, local models were a novelty. You could get them to run, but you wouldn't really use them for real agentic work. And that's changed, especially when Quen 3.5 came out. Things really changed. The CTO of Hugging Face said it was a before and after for local models, open source models, right? Quen 3.5 was that good. Is it as good as Claude code or Cursor on the hardest task? No. Is it slow? Yes. But for prototyping, learning quick scripts, working offline when you're flying on a plane, handling code you don't want to send to the cloud, this is genuinely usable now. Not works as a demo, but it is usable and it's free and it runs offline and it's only going to get better. The fact that a model you can run on a laptop is even in the same conversation as a $200 a month cloud tool or models is kind of wild when you think about where we were a year ago.
If you want to set this up yourself, I put the full guide on GitHub. Link in the description. Every command, the settings.json you can copy-paste, and the three demo prompts. I'll be linking that in the description. And if you're on 16 gigs RAM, don't worry. Quen 3.5 comes in sizes from 2 billion all the way to 122 billion. The 9 billion fits in the 16 gigs RAM. The 4 billion fits on 8 gigs. Subscribe if you want more of these videos. Thanks for watching and I'll see you.