Alibaba has made available as open source Qwen3-Coder, a novel and formidable large AI programming model belonging to the Tongyi Qianwen series. Leveraging a Mixture of Experts (MoE) architecture, this model boasts an impressive total of 480 billion parameters, with 35 billion of these being dynamically activatable. It is capable of handling contexts of up to 256K tokens and can be scaled to manage up to 1 million tokens. Qwen3-Coder underwent extensive pre-training on a vast dataset comprising 7.5 trillion data points, with 70% of this data being code-related. Through reinforcement learning specifically tailored for programming and agent tasks, its general capabilities, coding proficiency, and agent functionalities have been significantly enhanced. Across various evaluations, Qwen3-Coder's agent capabilities have surpassed those of GPT4.1, while its performance on the SWE-Bench benchmark stands in close competition with Claude4.