Qwen发布了其迄今最强大的开源代码模型Qwen3-Coder-480B-A35B-Instruct。该模型采用混合专家（MoE）架构，拥有4800亿参数，激活参数为350亿，支持高达256K的上下文长度，并可扩展至100万tokens。它精通358种编程语言，在SWE-bench-Verified等真实代码任务中表现出色，并具备自主与工具交互的agentic coding能力。此外，Qwen还同步开源了名为“Qwen Code”的命令行工具，以充分发挥模型潜力，旨在成为开发者生态系统的通用基础模型，推动自主软件开发新时代。

🌟 **强大的MoE架构与海量参数**：Qwen3-Coder-480B-A35B-Instruct采用了创新的混合专家（MoE）架构，拥有4800亿参数，但在推理时仅激活350亿参数，实现了高性能与高效率的平衡。这种设计使得模型在处理复杂任务时能够更具可扩展性和计算效率，为大型语言模型的发展开辟了新路径。

🚀 **超长上下文窗口与高可扩展性**：该模型原生支持256,000个token的上下文长度，并可通过Yarn等技术将其扩展至1,000,000个token。这意味着它可以一次性处理非常庞大的代码库、文档或项目，极大地提升了开发者在大型项目中的工作效率和代码理解能力。

💡 **卓越的Agentic Coding能力**：Qwen3-Coder被设计用于agentic coding工作流，不仅能够生成代码，还能自主与工具和开发环境进行交互，以解决复杂的编程问题。它在SWE-bench-Verified等真实世界代码任务基准测试中取得了领先的开源模型表现，甚至能与部分闭源模型相媲美，在代码重构、调试、自动化测试等方面展现出高超的熟练度。

🌐 **广泛的语言支持与生态整合**：模型支持358种编程和标记语言，涵盖了绝大多数主流开发语言。其作为“基础模型”的设计理念，使其能够轻松适应各种开发者工作流程、CI/CD管道和代码审查系统，为构建更广泛的AI驱动的开发生态系统奠定了基础。

🛠️ **配套开源工具“Qwen Code”**：为充分发挥Qwen3-Coder的能力，Qwen同步开源了名为“Qwen Code”的命令行agentic coding工具。该工具基于Gemini Code（gemini-cli）开发，并针对Qwen3-Coder进行了优化，支持自定义提示和高级函数调用协议，能够实现工具集成、多轮代码优化和上下文注入等agentic用例，方便开发者在终端直接调用模型进行代码开发。

Introduction

Qwen has unveiled Qwen3-Coder-480B-A35B-Instruct, their most powerful open agentic code model released to date. With a distinctive Mixture-of-Experts (MoE) architecture and comprehensive agentic coding capabilities, Qwen3-Coder not only sets a new standard for open-source coding models but also redefines what’s possible for large-scale, autonomous developer assistance.

Model Architecture and Specifications

Key Features

Model Size: 480 billion parameters (Mixture-of-Experts), with 35 billion active parameters during inference.Architecture: 160 experts, 8 activated per inference, enabling both efficiency and scalability.Layers: 62Attention Heads (GQA): 96 (Q), 8 (KV)Context Length: Natively supports 256,000 tokens; scales to 1,000,000 tokens using context extrapolation techniques.Supported Languages: 358 programming and markup languages, including Python, JavaScript, Java, C++, Go, Rust, and many more.Model Type: Causal Language Model, available in both base and instruct variants.

Mixture-of-Experts Design

The MoE approach activates only a subset of the model’s parameters for any given inference, delivering state-of-the-art performance with dramatically reduced computational overhead and enabling unprecedented scale.

Long Context and Scalability

Qwen3-Coder-480B-A35B-Instruct stands out for its native 256K context window, allowing direct handling of extremely large files and repositories. With context window extrapolation (using methods such as Yarn), it can scale up to 1 million tokens, making it suitable for even the largest codebases and documentation sets.

Performance Across Benchmarks

Agentic Coding

Qwen3-Coder is designed and optimized for agentic coding workflows—where the model not only generates code but autonomously interacts with tools and developer environments.

Benchmarks

SWE-bench-Verified: Achieves state-of-the-art results among open models on this challenging real-world coding task suite, outperforming or matching proprietary closed models in performance.Additional Agentic Tasks: Excels at Agentic Coding, Agentic Browser-Use, and Agentic Tool-Use, comparable to top-tier models such as Claude Sonnet-4.Breadth: Demonstrates high proficiency across competitive programming, automated testing, code refactoring, and debugging.

Foundation Model for Developer Ecosystems

Qwen3-Coder-480B-A35B-Instruct is built as a foundation model—intended to serve as a universal backbone for code understanding, generation, and agentic workflows across the digital world:

Maintains strengths in mathematics and reasoning, inherited from the Qwen3 base model.Adapts easily to various developer workflows, CI/CD pipelines, and code review systems.

Qwen Code: Open-Source Agentic Command-Line Tool

Overview

In tandem with the model, Qwen is also open-sourcing “Qwen Code”, a command-line agentic coding tool engineered to fully leverage the new model’s capabilities.

Key Features

Origin: Forked from Gemini Code (gemini-cli), ensuring compliance and open-source accessibility.Custom Prompts and Protocols: Enhanced with custom prompts and advanced function call protocols tailored for Qwen3-Coder, unlocking agentic use-cases such as tool integration, multi-turn code refinement, and context injection.Developer Integration: Designed to work seamlessly with best-in-class community tools, editors, and CI systems. Supports dynamic code interactions, repository-scale tasks, and direct function calling.Enhanced Tool Support: Utilizes an upgraded parser and function call logic to empower agentic workflows and program synthesis.

Usage and Extensibility

Qwen3-Coder-480B-A35B-Instruct is available under an open license and integrates with the broader open-source AI and development landscape. It can be run using standard transformers pipelines or through the dedicated Qwen Code CLI, and is compatible with modern developer stacks.

Conclusion

Qwen3-Coder-480B-A35B-Instruct marks a significant milestone in open-source code intelligence. With its blend of scalability, state-of-the-art agentic coding abilities, and developer-centric tooling, it provides a robust foundation model for the future of autonomous software development. Qwen’s commitment to openness—exemplified by both the release of the model and the Qwen Code agentic CLI—signals a new era for AI-powered, agentic coding in the global developer community.

FAQ 1: What are the main advantages of Qwen3-Coder-480B-A35B-Instruct compared to other open code models?

Qwen3-Coder-480B-A35B-Instruct stands out due to its massive scale—a 480B-parameter Mixture-of-Experts architecture with 35B active parameters—and its ability to natively handle 256,000-token contexts (scaling up to 1 million tokens via context extrapolation). This enables it to work with entire large codebases or repositories in one go. Its agentic design allows it not just to generate code, but also actively interact with developer tools and environments to autonomously solve complex programming tasks. Across multiple coding and agentic benchmarks, Qwen3-Coder delivers top-tier performance among open models, particularly excelling at SWE-bench-Verified and other real-world software engineering tasks.

FAQ 2: How do I use Qwen3-Coder with my own projects, and what is Qwen Code?

Qwen3-Coder-480B-A35B-Instruct can be accessed via standard Transformers pipelines or with the Qwen Code command-line interface, which is open-source and available on GitHub. Qwen Code, forked from Gemini Code, is a specialized agentic coding tool that leverages the model’s advanced custom prompts and function call protocols. It integrates easily with popular development stacks, supports seamless interaction with code bases and tools, and allows you to utilize Qwen3-Coder’s agentic capabilities for tasks such as code generation, refactoring, debugging, and automated tool use directly from your terminal.

FAQ 3: What kind of programming languages and tasks does Qwen3-Coder support?

Qwen3-Coder natively supports 358 programming and markup languages, including Python, JavaScript, Java, C++, Go, Rust, HTML, SQL, and many more. It is proficient at a wide spectrum of coding tasks, from competitive programming and code completion to bug fixing, code review, repository-scale understanding, test generation, refactoring, and multi-turn agentic workflows. Its long-context and foundation model architecture also make it suitable for integrating with CI/CD pipelines, cloud platforms, and large-scale software engineering environments.

---

Check out the Model on Hugging Face and Qwen Code GitHub Page. All credit for this research goes to the researchers of this project.

Join the fastest growing AI Dev Newsletter read by Devs and Researchers from NVIDIA, OpenAI, DeepMind, Meta, Microsoft, JP Morgan Chase, Amgen, Aflac, Wells Fargo and 100s more…….

The post Qwen Releases Qwen3-Coder-480B-A35B-Instruct: Its Most Powerful Open Agentic Code Model Yet appeared first on MarkTechPost.