XingChen-AGI introduced the Xing4.0-29B-A4B model

Developers from Xing4.0-29B-A4B have released a new MoE model with 29 billion parameters, optimized for Ascend NPU hardware. The release is focused on performing complex programming tasks and processing long sequences of data.

The model's architecture supports a context window of 256K tokens, allowing it to process significant amounts of code in one pass. In internal tests, the system demonstrated high performance, outpacing competitors in specialized benchmarks for developers, including SWE-bench Verified and Terminal-Bench 2.1.

The technical implementation of the model includes the use of a Mixture-of-Experts (MoE) architecture, which provides a balance between computational efficiency and response generation quality. Optimization for Ascend chips enables high inference speed in specific server environments where this infrastructure is deployed.

The model is already available for download in the repository on the Hugging Face platform. Developers note that the current version scores 223 points when performing test tasks in 2 hours, confirming the effectiveness of the chosen training strategy for automated programming tasks.

The release of Xing4.0-29B-A4B expands the range of tools for developers using NPU accelerators. The presence of long context support and specialized tuning for coding tasks makes the model suitable for integration into CI/CD pipelines and automatic code error correction systems.