Qwen3-Coder 30B-A3B Instruct

Qwen3-Coder 30B-A3B Instruct is a Mixture-of-Experts causal language model built for coding-focused workloads. With 30.5 billion total parameters but only 3.3 billion activated per forward pass across 128 experts, it balances strong capability with computational efficiency. The architecture features 48 transformer layers with grouped query attention and is optimized specifically for instruction-following without a "thinking mode" toggle. The model natively supports 262,144-token contexts, scalable to 1 million tokens via Yarn extension, making it suited for navigating large codebases and understanding multi-file repositories. It is engineered for agentic coding scenarios, including browser automation and function calling across platforms like Qwen Code and CLINE, using a specially designed format for tool interaction.

The model went through both pretraining and post-training stages, with emphasis on tool use and agentic workflows as supported by the catalog data. Released as an open-weight model under the Apache 2.0 license, it enables broad deployment flexibility for developers and teams. Performance analysis places it among the leading intelligence scores in its model class, though it operates slower than average and carries higher per-token costs. Its sweet spot lies in repository-scale code understanding, structured code completion, and seamless integration with OpenAI-compatible tool formats—making it particularly well-suited for automated coding agents, browser-based automation, and systems that orchestrate multiple external tools and functions.