Qwen3-Coder: Alibaba Qwen's Most Agentic Model for Code
Alibaba/Qwen
Alibaba Qwen has released Qwen3-Coder, a flagship code model with 480 billion total parameters and 35 billion active parameters, achieving state-of-the-art results among open models in agentic coding tasks, browser use, and tool use. The model supports 256K token context (up to 1M with extrapolation), was trained on 7.5 trillion tokens with 70% code, and uses large-scale reinforcement learning, including long-horizon RL for multi-step interaction in environments. Alongside the model, the Qwen Code tool based on Gemini Code has been open-sourced.
Alibaba Qwen announced Qwen3-Coder-480B-A35B-Instruct, a 480-billion parameter model based on Mixture-of-Experts architecture with 35 billion active parameters, setting a new record among open models in agentic coding tasks, browser interaction, and tool use, rivaling Claude Sonnet 4. The model natively supports a 256K token context, extendable to 1 million tokens via extrapolation methods (YaRN). Qwen3-Coder was trained on 7.5 trillion tokens (70% code) while retaining general and mathematical abilities. For post-training, two large-scale approaches were used: Code RL (reinforcement learning on diverse coding tasks with automatic test case generation) and Long-Horizon RL for multi-step interaction in real-world software engineering tasks, such as SWE-Bench. To scale the training environment, Alibaba Cloud infrastructure was used, capable of running 20,000 independent environments in parallel. A command-line tool, Qwen Code, based on Gemini Code with adapted prompts and function-calling protocols, is released alongside the model. Qwen3-Coder can also be used with Claude Code, Cline, and via the Alibaba Cloud Model Studio API. The authors plan to release other model sizes and explore self-improvement for coding agents.
Source: Alibaba Qwen —
original