XingChen-AGI released the MoE model Xing4.0 with 29 billion parameters

Developers from XingChen-AGI introduced a new neural network architecture Xing4.0-29B-A4B. The model is built on the basis of the MoE (Mixture of Experts) architecture and is designed to perform complex agent tasks. The total number of model parameters is 29 billion, with 4 billion active parameters involved in calculations. This configuration allows optimizing the consumption of computational resources while maintaining high performance. The model supports an extended context window of 256K tokens. A quantized version in GGUF format is available for local launch and integration into workflows, with a size of 18 GB. The architecture has been specially optimized for work on Ascend NPU hardware accelerators, expanding its application possibilities in specialized server environments. In preliminary tests, the model showed a result of 107 points, achieving this score in 4 hours of work. The release is aimed at developers creating autonomous AI agents that require high accuracy when processing long data sequences. The release of Xing4.0 continues the trend of creating specialized MoE solutions that allow effectively balancing between response speed and generation quality. The use of GGUF quantization makes the model available for deployment on consumer and corporate equipment with limited video memory.