Qwen3-Coder is the most agentic code model to date in the Qwen series.

Overview

qwen3-coder:30b offers 30B total parameters with only 3.3B activated, delivering strong performance while maintaining efficiency.

• Exceptional agentic capabilities for real-world software engineering tasks through advanced long-horizon reinforcement learning on SWE-Bench and similar benchmarks.
• Long context support with 256K tokens natively and up to 1M tokens using extrapolation methods, optimized for repository-scale understanding.
• Scaled pretraining on 7.5T tokens (70% code ratio) while preserving strong general and mathematical abilities.
• Execution-driven reinforcement learning that significantly boosts code execution success rates across diverse real-world coding tasks.

Reference

• Blog