Abstract: The review of Environmental Impact Assessment(EIA) reports is a core component of environmental regulatory oversight, yet traditional auditing processes rely heavily on manual labor and suffer from low efficiency. Leveraging the open-source large language model DeepSeek as a technological foundation, this study proposes a “multimodal knowledge fusion-dynamic risk control-human-machine collaborative iteration” tripartite intelligent auditing framework, achieving a paradigm shift in EIA review methodology. Empirical results demonstrate that DeepSeek reduces the error rate in pollutant accounting from 7.2%(manual audit) to 0.8%, enhances the accuracy of logical flaw identification compared to existing tools, and shortens the audit time per report to the minute level. This research provides a theoretical framework and practical pathway for AI-driven environmental governance, supporting the strategic goal of “intelligent transformation” outlined in the 14 th Five-Year Plan for the Reform of Environmental Impact Assessment and Pollution Discharge Permit Management.