What changes in the new release
Alibaba has announced version 1.0 of Qwen Code, presented as the result of merging Qoder and Qwen Code into a single open-source core. According to the source, the company has transferred the Harness capabilities developed in Qoder into Qwen Code, making them available together with Qwen Code 1.0 Preview. The stated goal is to offer developers an execution foundation for agents that can be embedded in their own products.
The Harness component is described as the layer that connects model, tools, and execution process. It is at this junction that developers can build agent systems suited to their own scenario and insert AI-assisted programming capabilities into more elaborate workflows. The novelty, therefore, concerns not a single model but the infrastructure that organizes its operational use.
Architecture: from in-memory events to distributed nodes
The new system starts from an in-memory event bus within the process and moves toward a distributed model. The transition goes through a layering of access, planning, and execution, with a separation defined by the source as the distinction between reasoning and action: inference orchestration and tool execution are managed independently. The stated shift is from a single process with multiple sessions to multiple nodes with a larger number of sessions, while retaining existing inputs and response modes.
Alibaba maintains that the uniform event mechanism not only connects inputs and tasks but also defines clear boundaries for capability extension and resource management. The various modules operate around events, while the model, sandbox, and external extensions are administered separately. According to the company, this results in greater ease of extension, centralized control, and cost management when the system must support multiple tasks.
Centralized control and resource management
The material lists four directions. Processing units can add inference, memory, and tool-use capabilities based on events, reducing intrusiveness on the main flow. Model access goes through a single entry point, with configuration, authorizations, and call logs managed centrally. Sandbox resources are allocated and reused within the same configuration and permission limits, to contain repeated initializations and idle usage. Finally, configuration, authorization, and MCP connections are administered in a unified way.
Availability and open questions
The Preview version is distributed on a separate branch, while the existing stable branch remains available; the team indicates its intention to continue along the 1.0 trajectory by improving functional coverage and stability. The source also reports the URL of the project's GitHub repository. The actual performance of the distributed architecture, the maturity level of the Preview branch, and the licensing and governance conditions of the open-source core remain to be independently verified—aspects that the announcement does not detail.
