Back to Blog

Qwen3-Coder: Agentic Coding in the World

Qwen Image Team•August 12, 2025•3 min

We are proud to introduce Qwen3-Coder, our most advanced **agentic code model** to date. Today, we’re releasing its flagship variant — **Qwen3-Coder-480B-A35B-Instruct** — a **480B-parameter Mixture-of-Experts** model with **35B active parameters**, delivering **state-of-the-art performance** in both **coding** and **agentic reasoning**.

Qwen3-Coder: Agentic Coding in the World

We are proud to introduce Qwen3-Coder, our most advanced agentic code model to date.
Today, we’re releasing its flagship variant — Qwen3-Coder-480B-A35B-Instruct — a 480B-parameter Mixture-of-Experts model with 35B active parameters, delivering state-of-the-art performance in both coding and agentic reasoning.

With native 256K context length and support for 1M tokens via extrapolation methods, Qwen3-Coder is built to handle large-scale, complex coding tasks. It sets new benchmarks among open models for Agentic Coding, Agentic Browser-Use, and Agentic Tool-Use, performing on par with Claude Sonnet 4.

Key Highlights

  • Massive Scale & MoE Architecture

    • 480B parameters, 35B active per query.
    • Mixture-of-Experts design for efficiency and flexibility.

  • Unmatched Context Handling

    • Native 256K context support.
    • Extendable to 1M tokens with YaRN, enabling repo-scale comprehension.

  • State-of-the-Art Benchmarks

    • Leading results on Agentic Coding, Agentic Browser-Use, Agentic Tool-Use, and SWE-Bench Verified.

Open-Source CLI: Qwen Code

Alongside the model, we are releasing Qwen Code — an open-source command-line tool for agentic coding, forked from Gemini Code and tailored with custom prompts and function-calling protocols to unleash Qwen3-Coder’s full potential.

Qwen3-Coder integrates seamlessly with the community’s top developer tools, empowering developers to build, debug, and ship code with intelligent assistance.

Pre-Training Advances

We pushed the limits of pretraining in three major ways:

  1. Scaling Tokens

    • Trained on 7.5T tokens with a 70% code ratio, excelling in code generation while retaining strong general and math capabilities.

  2. Scaling Context

    • Optimized for repo-scale understanding and dynamic data such as Pull Requests.
    • Natively supports 256K context, extendable to 1M.

  3. Scaling Synthetic Data

    • Leveraged Qwen2.5-Coder for data cleaning and rewriting, drastically improving data quality.

Post-Training Innovations

1. Scaling Code RL: Hard to Solve, Easy to Verify

Instead of focusing solely on competitive code generation, we scaled execution-driven reinforcement learning across real-world coding tasks. By automatically generating diverse test cases, we created high-quality training instances that significantly improved code execution success rates and boosted performance on other tasks.

2. Scaling Long-Horizon RL for Agentic Tasks

Real-world engineering tasks like SWE-Bench require multi-turn interaction — planning, tool use, feedback loops, and decision-making.

In post-training, we introduced Agent RL to train Qwen3-Coder for long-horizon, tool-assisted problem solving.
To support this, we built a scalable system capable of running 20,000 parallel environments on Alibaba Cloud, providing real-time feedback for large-scale RL.

As a result, Qwen3-Coder achieves state-of-the-art performance on SWE-Bench Verified among open-source models — without the need for test-time scaling.

A Foundation for Agentic Coding Everywhere

Qwen3-Coder is designed to power Agentic Coding in the World — from IDEs to cloud-based development platforms, from automated software agents to large-scale code analysis pipelines.

We invite the community to try Qwen3-Coder, contribute, and help shape the future of agentic, intelligent coding.

👉 Explore more at: https://www.qwenimagen.com/