Qwen3-Coder
https://qwenlm.github.io/blog/qwen3-coder/
The Chinese Qwen team, behind the development of advanced AI models, announced the release of Qwen3-Coder. The Qwen3-Coder-480B-A35B-Instruct model uses a Mixture-of-Experts architecture with 480 billion parameters (of which 35 billion are active), supports a context window of up to 256k tokens out-of-the-box, and can be extended to 1 million tokens. Other sizes are expected to be released.

During the post-training phase, the Qwen team scaled up reinforcement learning for code (Code RL), focusing on real-world tasks where execution success is easily verifiable. Additionally, they introduced Long-Horizon Reinforcement Learning (Long-Horizon RL or Agent RL) to teach the model to solve complex engineering problems, such as SWE-Bench, through multi-step interaction with the environment, including planning, tool use, and feedback acquisition.

The model can integrate with Claude Code and Cline.

https://qwenlm.github.io/blog/qwen3-coder/
For interaction with Qwen3-Coder, the developers introduced a command-line tool CLI – Qwen Code, which is essentially a Chinese copy of Gemini Code.

We get performance at the level of Claude 4 Sonnet, only significantly cheaper.
https://openrouter.ai/qwen/qwen3-coder

#qwen