Qwen Launches Qwen3-Coder-480B-A35B-Instruct: The Most Advanced Open Agentic Coding Model

Revolutionary Model Architecture

Qwen has introduced the Qwen3-Coder-480B-A35B-Instruct, a groundbreaking open agentic coding model boasting 480 billion parameters with a Mixture-of-Experts (MoE) architecture. This model activates 35 billion parameters during inference through 160 experts, with 8 experts active per inference, balancing efficiency and scalability. It features 62 layers and sophisticated attention heads, supporting an unprecedented native context length of 256,000 tokens, extendable up to 1 million tokens using context extrapolation techniques.

Extensive Language and Task Support

The model supports 358 programming and markup languages, including popular ones like Python, JavaScript, Java, C++, Go, and Rust. As a causal language model, it comes in both base and instruct variants, optimized specifically for agentic coding workflows where the model autonomously interacts with developer tools and environments.

Performance and Benchmarks

Qwen3-Coder-480B-A35B-Instruct achieves state-of-the-art results on the SWE-bench-Verified benchmark, outperforming or matching proprietary closed models. It excels in agentic coding, browser use, and tool use, on par with leading models such as Claude Sonnet-4. The model demonstrates high proficiency in competitive programming, automated testing, code refactoring, and debugging.

Foundation Model for Developer Ecosystems

Designed as a foundation model, Qwen3-Coder integrates strengths in mathematics and reasoning from its base model and adapts seamlessly to various developer workflows, including CI/CD pipelines and code reviews. It serves as a universal backbone for code understanding, generation, and autonomous developer assistance.

Open-Source Agentic CLI: Qwen Code

Alongside the model, Qwen has open-sourced Qwen Code, a command-line agentic coding tool forked from Gemini Code (gemini-cli). This tool leverages custom prompts and advanced function call protocols tailored for Qwen3-Coder, enabling multi-turn code refinement, tool integration, and context injection. It supports dynamic code interactions, repository-scale tasks, and direct function calling, integrating smoothly with popular editors and CI systems.

Accessibility and Integration

Qwen3-Coder-480B-A35B-Instruct is available under an open license and can be utilized through standard transformers pipelines or the Qwen Code CLI. It aligns with modern developer stacks and supports extensive agentic workflows, marking a significant advancement in open-source code intelligence and autonomous software development.

FAQs

Main advantages: Massive scale with Mixture-of-Experts architecture, large native context window (256K tokens), agentic design for autonomous interaction with developer tools, and leading benchmark performance.

Usage: Accessible via Transformers pipelines or Qwen Code CLI, which supports deep integration with developer tools and workflows.

Supported languages and tasks: Supports 358 languages and a wide range of coding tasks, including competitive programming, debugging, refactoring, and CI/CD integration.

For more information, visit the Model on Hugging Face and the Qwen Code GitHub page.