Qwen3-Coder: An Open-Weight Coding Model and Its Command-Line Agent

요약

30초 요약

Matthew Berman reviews Qwen3-Coder-480B-A35B-Instruct and Qwen Code, describing their architecture, long context, training and reported agentic coding benchmarks. He shows coding demonstrations and builds a Snake game, which works with some lag.

1A narrow benchmark comparisonQwen’s chart reports 69.6% for Qwen3-Coder and 70.4% for Claude Sonnet 4 on SWE-bench Verified with OpenHands and 500 turns. The result is specific to the stated evaluation setup.
2Mixture of experts and long contextThe 480B model activates 35B parameters and supports 256K tokens natively, with an advertised extension to one million using YaRN.
3Code-heavy trainingQwen reports 7.5 trillion pretraining tokens with a 70% code ratio and says it used Qwen2.5-Coder to clean noisy data.
4Long-horizon reinforcement learningQwen says its infrastructure could run 20,000 independent environments in parallel to train and evaluate tool-using agents.
5A quick Snake testBerman reports that his game took about a minute and 792 lines to generate. It ran, although he noticed some lag.

핵심 정리

Qwen3-Coder’s released model has 480 billion total parameters, 35 billion active parameters and a 256K native context window.
Qwen reported 69.6% on SWE-bench Verified with OpenHands and 500 turns, close to the 70.4% figure it reported for Claude Sonnet 4 under that configuration.
Qwen describes code-focused pretraining and reinforcement learning with up to 20,000 parallel environments.
Qwen Code is a command-line coding agent adapted from Gemini CLI.
Berman’s single Snake game demo worked with some lag; it is an illustration rather than a comprehensive evaluation.

주제코딩 에이전트 Open-weight models 모델과 학습 Benchmarks Reinforcement Learning

전문

챕터별로 다듬은 전문 읽기

원본 자막과 대조

원본 자막만 보기

Benchmark and model architecture

Berman compares Qwen’s reported SWE-bench result with Claude Sonnet 4. He introduces the mixture-of-experts architecture, long context and Qwen Code.

Qwen has released Qwen3-Coder, an open-weight coding model that I think belongs in the same conversation as Anthropic’s Claude models. The Qwen team’s SWE-bench Verified chart reports 69.6% for Qwen3-Coder and 70.4% for Claude Sonnet 4 using OpenHands with 500 turns. That is a close result on this particular test and configuration, not proof that the models are interchangeable across every coding task.

슬라이드 1 / 4Qwen’s SWE-bench Verified comparisonThe chart displays Qwen’s reported model results, including the 500-turn OpenHands comparison.

Qwen also released a command-line coding tool, Qwen Code, adapted from Gemini CLI with custom prompts and function-calling support. The first and largest model in the July 2025 release is Qwen3-Coder-480B-A35B-Instruct: a mixture-of-experts model with 480 billion total parameters and 35 billion active parameters. Its model card lists an Apache 2.0 license.

The model supports a 256,000-token context window natively, and Qwen says YaRN can extend it to one million tokens. A long context can be useful for coding because an agent may need to work across a repository, tool outputs and changing information. Qwen positions the model for coding, tool calls and other agentic tasks.

Pretraining and agent reinforcement learning

The Qwen team describes code-heavy pretraining, data cleaning and execution-driven reinforcement learning. Its long-horizon training infrastructure supports parallel environments.

슬라이드 2 / 4Pretraining and code reinforcement learningQwen lists the training-token mix, context extension and synthetic-data cleaning approach.

The team’s starting premise is that there is still room to scale pretraining. Qwen reports 7.5 trillion training tokens with a 70% code ratio, and says the model retains general and mathematical abilities. It also used the earlier Qwen2.5-Coder model to clean and rewrite noisy data, so one generation helped prepare training material for the next.

For post-training, Qwen describes scaling code reinforcement learning across a broader range of real-world tasks rather than focusing only on competitive programming. It generated diverse test cases to build training instances, reporting improvements in code execution and other tasks. The attraction is that a coding problem can be hard to solve while its proposed solution is comparatively easy to verify by execution.

Long-horizon agent reinforcement learning adds multi-turn interactions: planning, using tools, receiving feedback and deciding what to do next. Qwen says it built infrastructure on Alibaba Cloud capable of running 20,000 independent environments in parallel for training and evaluation. The company reports its SWE-bench result without test-time scaling; that statement alone does not establish that the model cannot reason, and my speculation about future improvements remains speculation.

Trying the model and viewing demos

Berman shows a hosted coding interface and examples of generated interactive applications. The examples include a physics simulation and several visualizations.

At the time of the video, Hugging Face hosted a way to try the model. In the example I choose a bouncing-ball task: the interface writes code and can execute it there. I then show demonstrations made with Qwen Code, starting with a physics simulation of a structure breaking apart.

슬라이드 3 / 4Physics simulation demonstrationAn example shows a structure breaking into blocks.

Other examples include an interactive color visualization, a 3D Earth terrain simulation, a typing-speed app, a bouncing ball and rotating hypercube, a solar system and a game called Duet. These clips show a range of applications the tool can generate, though a selected demo does not by itself measure reliability on a broader set of projects.

A Snake game test

Berman uses Qwen Code to generate a complex Snake game. He finds it functional but somewhat laggy.

Finally, I load Qwen Code in my command line and ask it to make a complex Snake game. It takes about 60 seconds and produces 792 lines of code. The game runs when I open it; I am not sure what one object at the top is doing, and it is a little laggy, but it is still a promising quick test.