DeepSeek announced the release of its V4.1 Flash AI model on Thursday, positioning the update as a significant advancement in the ongoing competition within China’s artificial intelligence sector. The company reports that the new model outperforms the Kimi K3, developed by Moonshot AI, specifically in benchmarks related to coding and cybersecurity tasks.

The technical foundation of the V4.1 Flash model is a new architecture described as a "Causal-Encoder-Decoder." Despite being built on a framework containing 552 billion parameters, the model employs a Mixture-of-Experts (MoE) design. This approach allows the system to be more efficient than traditional models, which typically process every query through the entire network. By activating only relevant parts of the model for specific tasks, DeepSeek claims to have achieved both increased processing speeds and reduced inference costs.

This release is part of a broader trend among Chinese AI firms, which are currently engaged in an aggressive competition focused on both price reductions and performance benchmarks. While DeepSeek has publicly touted these performance gains, the claims are based on internal benchmarks provided by the company. Independent verification of these performance metrics against the Kimi K3 has not yet been provided by third-party analysts.