XingChen-AGI released the quantized model Xing4.0-29B-A4B-FP8

Developers from XingChen-AGI presented an updated version of the Xing4.0-29B-A4B-FP8 neural network. The model represents a MoE (Mixture of Experts) type architecture with a total of 29 billion parameters, of which 4 billion are active when executing a request.

The key feature of the release was optimization for Ascend NPU hardware. The model supports working with a context window of up to 256 thousand tokens, which allows processing significant amounts of data within a single session. During internal tests, the system scored 62 points, spending 3 hours on calculations.

The architecture is focused on performing complex agent tasks and programming. The model has been tested on specialized benchmarks, including SWE-bench Verified, which evaluates the ability of AI to solve real problems in open-source repositories. The use of the FP8 format made it possible to reduce memory requirements while maintaining performance.

This release expands the possibilities for deploying high-performance models on specialized computing accelerators. The use of FP8 quantization provides a balance between inference speed and answer accuracy, which is critical for automated systems development and code analysis.