Qwen 3 Coder in 6 Minutes

--- type: transcript date: 2025-07-24 youtube_id: gqzsFWZe0Iw --- # Transcript: Qwen 3 Coder in 6 Minutes In this video, I'm going to be going over Quen 3 Coder, which is the latest model from the team over at Alibaba. This is their most powerful open age code model to date. This is a 480 billion parameter mixture of experts model with 35 billion active parameters as well as a context length that scales all the way to a million tokens of contexts with extrapolation. Now, what's really impressive with this model is we just had Kimmy K2 that really excelled at a ton of different coding tasks. But what's really impressive with this latest release from Quen is we can see across the board outperformance from the Kimmy K2 model. Where this is really interesting is we can actually see Quen 3 outperform even Claude for Sonnet across a number of different benchmarks. And where it doesn't outperform, it does come quite close. We can see on terminal bench it outperforms. For SWEBench verified, we can see it comes in at 69.6, whereas Claude 4 comes in at 70.4. And across a number of different benchmarks like Agentic browser use, we can see that this is almost a tossup in terms of the comparison to Cloud 4 sonnet. And then for a Gentic tool use, while it doesn't quite meet Cloud for Sonnet performance, it does come within striking distance basically across the board. So now in terms of the command line tool, they're open sourcing Quen Code, which was forked from Gemini code. And Quen code has been adapted with customized prompts and function calling protocols to fully unleash the capabilities of Quen 3 coder on a gentic coding task. And furthermore, if you're interested in using Quen 3 coder in something like Klein, you are going to be able to leverage that. Or additionally, if you have an AI IDE that does support being able to swap in the base URL as well as the model string and point to different model providers, you will be able to try out Quen 3 coder within whatever platform that supports it. Now in terms of the specifics, so we've definitely seen a lot of focus on these hybrid reasoning models as well as a lot of the hype around testime compute. But what's interesting with Quen 3 coder as well as coding models in general is speed is definitely a very important factor especially when leveraging these things within something like an AI IDE or increasingly within these agentic tools like whether it's cloud code or Gemini CLI or now this Quen CLI that they just put out. In terms of some of the specifics, so the model was trained on 7.5 trillion tokens, 70% of which was actually coding tokens. Also, they leveraged synthetic data to actually filter out a lot of the noisy data, which they mentioned that significantly improved the overall data quality. It natively supports up to 256,000 tokens, but it can be extended up to a million tokens with yarn. And this was optimized for repo scale and dynamic data, things like pull requests to empower things like agentic coding tools. Another aspect, there isn't a prevailing focus on competition level code generation. They really just focused on tasks that were well suited for execution-driven large-scale reinforcement learning. And then they scaled up the code RL training on a broad set of realw world coding tasks. So now just to quickly put this into perspective for the benchmarks, we can see it comes within spitting distance of claw force sonnet and we can see Kimmy K2 that just came out at 65.4. And now what's really impressive is if we think back to Deep Seek R1 when that came out. So just a number of months ago, everyone was talking about them. Everyone was very impressed with their benchmarks. Here we see an absolute huge increase in just a number of months. I think this is just further proof that things are definitely not slowing down. If anything, there is definitely a case to be made that things are continuing to accelerate. Another interesting part of the blog post, they mentioned that for software engineering tasks like Swebench, Quen 3 coder must engage in multi-turn interactions with the environment, things like planning as well as tool use. They mentioned in post training for Quen 3 coder, they introduce long horizon reinforcement learning to encourage the model to solve real world problems through multi-turn interaction with tools. The key challenge of agent RL lies within the environment scaling. They describe what they did is leverage 20,000 independent environments in parallel leveraging Alibaba's cloud infrastructure, the infrastructure that was necessary to provide feedback for the large-scale reinforcement systems and support the evaluations at scale. In other words, it looks like they've spun up these different instances to actually see how these different agents would perform. And as a result, this is what allowed Quen 3 coder to achieve state-of-the-art performance across open source models. So to get started with our CLI tool, it is straightforward. We can go ahead and we can mpm install Quen Code and then we'll just have to grab our API keys from whether it's Open Router or whoever really is hosting it. We just have to make sure that we swap out the OpenAI API key, the base URL as well as the model to the relevant Quen model. Now, additionally, if you do want to leverage this within Cloud Code, I saw a lot of people try this out when Kimmy came out. All that we need to do this is we can get an API key from the Alibaba Claude model studio. We can install cloud code and then we can go ahead and set the proxy URL as well as the O token. Additionally, client is another option that a lot of people are fans of. So, you can go ahead and try it within there. Now, in terms of some demonstrations, here is a demonstration of what looks like a web app, maybe 3JS, where it has some realistic physics of a building or a bunch of blocks falling down. And then further, if we go through some of these other examples, we can see this Earth visualization. Again, this might be something like 3JS that generated that. We can also see these little web apps that we've seen a lot of these LLMs be able to generate. Additionally, here's the bouncing ball in a rotating hyper cube. Additionally, we have a solar system simulation as well as a little game by the looks of it. And finally, if you do want to try this out within a web app interface, you can head on over to chat.quendi and you'll be able to try this out completely for free on their platform. And what is nice with their interface is they do also have an artifacts feature where you can see little web apps of whatever you ask it to generate. By the looks of it, the imprint speed on the platform definitely is pretty reasonable. So, you will be able to try this out and get a general sense on the models capability within here. Otherwise, that's it for this video. If you found this video useful, please comment, share, and subscribe.