Qwen Team introduced Qwen3-Coder, a family of advanced code models built for complex programming tasks requiring long contexts. The leading model, Qwen3-Coder-480B-A35B-Instruct, utilizes 480 billion parameters with a focus on execution, decision-making, and adaptability. It supports native handling of 256K tokens and extends to 1 million for large inputs. This model incorporates post-training reinforcement learning, ensuring generated code functions correctly. Qwen also developed Qwen Code, an open-source command-line interface that facilitates enhanced coding support for various programming environments.
Qwen3-Coder is a new family of agentic code models designed for long-context, multi-step programming tasks, with significant advancements in handling complex programming challenges.
The most capable variant, Qwen3-Coder-480B-A35B-Instruct, is designed with 480 billion parameters and is optimized for execution and decision-making in code generation.
Qwen3-Coder emphasizes a post-training reinforcement learning approach, focusing on the success of generated code based on its ability to run and solve problems.
Qwen Code, an open-source command-line interface, is available for developers, enhancing tool support, function calling, and compatibility with multiple environments.
Collection
[
|
...
]