Wenle Yu Yang from Aofei Temple

Quantum Bit | Public Account QbitAI

The Throne of the Programming Agent, Taken by Domestic Open-Source Models!

Just now, the Tongyi large model team from Alibaba open-sourced Qwen3-Coder, directly setting a new SOTA for AI programming —

Not only did it surpass DeepSeek V3 and Kimi K2 in the open-source community, but it also outperformed the industry benchmark, the closed-source Claude Sonnet 4.

Netizens immediately tested it with a small ball bounce, and the result was like this:

The effect was so strong that people were amazed: It's almost changing the rules of the game.

After all, this is open-source and free!

Now everyone doesn't have to spend $200 a month on Claude Code anymore!

Qwen3-Coder comes in multiple sizes, with the strongest version being Qwen3-Coder-480B-A35B-Instruct, a 450B MoE model with an activated parameter count of 35B.

It natively supports a 256K context length and can be extended to 1M using YaRN.

The command-line version of Qwen also made its debut:

The Tongyi team adapted prompts and tool call protocols based on Gemini Code, and then developed and open-sourced the command-line tool Qwen Code.

Well, these days, if you don't have a CLI, it's hard to say you're a programming agent (doge).

Amazing Results from Simple Prompts

How does Qwen3-Coder perform? Let's just see for ourselves.

Generally, you can get surprising results with simple language:

For example, one sentence can create a colorful, interactive animation using p5js.

A 3D Earth visualization, and you can get a digital globe in a minute.

You can also create dynamic weather cards.

Interactive mini-games are also easy to handle.

We also did a quick test, starting with the most practical feature — making a resume.

The prompt was: Generate an editable resume template.

Creating a Minesweeper game is also a breeze, write and play right away~

The prompt was: Generate a Minesweeper game.

Beyond the effects, what's worth noting is that this time, the Tongyi team also released many technical details from pre-training to post-training.

Technical Details

In the pre-training phase, Qwen3-Coder mainly focused on scaling from different angles to enhance the model's capabilities.

This includes data expansion, context expansion, and synthetic data expansion.

The training data size reached 7.5T tokens, with 70% being code data. This ensures general and mathematical abilities while improving programming skills.

At the same time, it natively supports a 256K context length, and with the YaRN technology, it can be extended to 1M, suitable for repository-level and dynamic data processing.

In training, Qwen2.5-Coder was used to clean and rewrite low-quality data, significantly improving the overall data quality.

Different from current models that focus on competitive programming, the Qwen team believes that code tasks are naturally suitable for large-scale reinforcement learning driven by execution.

Therefore, in the post-training phase, they expanded the training on rich and real code tasks through Scaling Code RL, automatically generating diverse test cases, which improved the success rate of code execution.

On the other hand, they introduced Scaling Long-Horizon RL, building a system capable of running 20,000 independent environments simultaneously using Alibaba Cloud infrastructure, allowing the model to perform well in multi-turn interactions, especially achieving SOTA results on SWE-bench Verified among open-source models.

Open Source vs Closed Source

By now, are you also eager to try the real power of Qwen3-Coder?

We'll guide you:

  • The simplest way is to experience it directly on the Qwen official website;
  • Install Qwen Code via the command line, supporting OpenAI SDK to call LLM;
  • Apply for an API on the Alibaba Cloud BaiLian platform, and you can use it together with programming tools like Claude Code and Cline.

It's worth mentioning that Qwen3-Coder still follows the Apache License Version 2.0, which is friendly for commercial use.

After all, since it's open source, the initiative has already been handed over to developers ~

This is also the key reason why Qwen's release caused such a surge of online enthusiasm:

Qwen3-Coder seems to be a major leap forward for open-source programming agents.

Is it now equal? Will surpassing not be far behind?

What's even more exciting is that, on the path of open source, Chinese models are undoubtedly leading the way.

Official website:

https://chat.qwen.ai/

Project address:

https://github.com/QwenLM/qwen-code

Reference link:

https://mp.weixin.qq.com/s/CArpTOknOQC5O90Wgih3SA

— End —

Quantum Bit QbitAI

Follow us to stay updated on the latest technological developments.

Original article: https://www.toutiao.com/article/7530067412733936167/

Statement: The article represents the personal views of the author. Please express your opinion by clicking on the [Up/Down] buttons below.