On Mar. 17, Tsinghua University and startup Qingcheng.AI announced the release of an AI framework named Chitu.

This framework aims to reduce dependency on Nvidia GPUs for large language model (LLM) inference tasks.

Chitu supports Meta’s Llama series and DeepSeek models, running on China-made chips and challenging Nvidia’s dominance.

Tests on Nvidia’s A800 GPUs showed Chitu increased inference speed by 315% and cut GPU usage by 50% when running DeepSeek-R1.

China’s AI firms are actively seeking alternatives to Nvidia, driven by US restrictions on advanced GPU exports.