Qwen Releases 480B Open-Source Model for Code Agents

Alibaba's Qwen team has released a powerful new open-source model specialized for code generation and agentic tasks. The model, named Qwen3-Coder-480B-A35B-Instruct, marks a significant entry into the high-end landscape of AI coding assistants, combining immense scale with an efficient architecture.

At its core, Qwen3-Coder is a Mixture-of-Experts (MoE) model. While it contains a massive 480 billion total parameters, it only activates a sparse 35 billion for any given input. This design aims to provide the vast knowledge capacity of a very large model while maintaining more manageable computational requirements for inference, striking a balance between power and efficiency.

Long Context and a Permissive License

Beyond its MoE architecture, the model boasts two key features that set it apart. First is its exceptionally long context window of 262,144 tokens, allowing it to process and reason over entire codebases or lengthy project histories. This is critical for complex, multi-file software development tasks.

Second, and perhaps most importantly for the community, Qwen3-Coder is released under the Apache-2.0 license. This permissive license allows for both commercial and research use, giving developers a powerful, unrestricted foundation to build sophisticated new tools and coding agents. You can find the model and its details on its Hugging Face repository.