Qwen Releases Compact 30B MoE for Coding Agents

Alibaba's Qwen team has released Qwen3-Coder-30B-A3B-Instruct, a new open-source model specialized for code generation and agentic tasks. Published under the permissive Apache 2.0 license, the model is designed to offer a balance of high performance and computational efficiency for developers building AI-powered tools.

The model's architecture is a key feature. It is a Mixture-of-Experts (MoE) model with a total of 30 billion parameters, but it only activates 3 billion of them during any given inference pass. This design allows it to access a large repository of knowledge while keeping the computational cost relatively low, making it more accessible to run on consumer-grade hardware compared to dense models of a similar size.

Key Specifications

• Total Parameters: 30 billion
• Active Parameters: 3 billion (MoE)
• Context Length: 262,144 tokens
• License: Apache 2.0

The combination of an efficient MoE architecture and a very long 262k token context window makes Qwen3-Coder particularly well-suited for building autonomous coding agents. These agents require the ability to understand large codebases and maintain context over complex, multi-step tasks. You can find the model and further details on its Hugging Face repository page. Its release provides another strong, permissively licensed option for teams looking to integrate advanced code generation into their workflows.