Alibaba unveils Qwen3-Coder with support for AI agents capable of making independent decisions
The Qwen team has officially presented Qwen3-Coder, its most powerful code generator to date, which supports the "agent" approach. The main star of the line is Qwen3-Coder-480B-A35B-Instruct. This is a hybrid Mixture-of-Experts model with 480 billion parameters (35 billion active at the same time) that can work with the context of up to 256,000 tokens natively and up to 1 million tokens with extrapolation. In tests, it demonstrates record results among open source models in Agentic Coding, Agentic Browser-Use, and Agentic Tool-Use tasks - at the Claude Sonnet 4 level.
What is agentic coding?
Qwen3-Coder was created not only to generate code, but also to work as an "agent" that can plan, use tools, receive feedback, and make decisions. In real-world tasks (e.g., SWE-Bench), it interacts with the environment in several steps, which requires complex Long-Horizon RL learning. For this purpose, the Qwen team has raised a scalable system of 20,000 parallel environments based on Alibaba Cloud.
Qwen3-Coder features
- Trained on 7.5 trillion tokens, where 70% is code
- Supports giant context for up to 1 million tokens (with YaRN)
- Using reinforcement learning (Code RL) on real-world tasks to improve code execution
- Integration with developer tools such as Qwen Code (CLI), Claude Code, and even OpenAI SDK
- Ability to work both through the API and in the local environment
Tools and APIs
The Qwen Code CLI tool based on Gemini Code is available with the model. It allows you to easily use Qwen3-Coder in your own projects, with support for custom prompts and functions. It also provides work through Claude Code and direct API access through Alibaba Cloud Model Studio.
What is it for?
Qwen3-Coder is positioned as a universal coding agent - from website generation to complex multi-step development tasks. The Qwen team plans to release more model sizes to reduce deployment costs and is exploring opportunities for agent self-improvement.
About Qwen
Qwen is a series of large-scale language models and agent solutions from Alibaba's Qwen Team, designed to generate text, code, data analysis, and multi-step scenarios using tools. The latest version, Qwen3-Coder, is built on a Mixture-of-Experts architecture with 480 billion parameters (35 billion active), supports up to 256,000 context tokens natively and up to 1 million with extrapolation, and demonstrates closed model-level results like Claude Sonnet 4 in SWE-Bench, Agentic Coding, Browser-Use, and Tool-Use tasks. The model integrates with the OpenAI SDK, CLI tools (Qwen Code, Claude Code), Alibaba Cloud API, and supports GitHub and Hugging Face ecosystems, using reinforcement learning on real-world coding tasks to improve accuracy and efficiency.
