Google Research Launches Gemini-SQL2, Beating Prior Single-Model BIRD Mark by 2.5%
Google Research introduced Gemini-SQL2, a text-to-SQL capability powered by Gemini 3.1 Pro that translates natural-language prompts into execution-ready SQL queries. The company said the system achieved state-of-the-art results on the BIRD benchmark, with a 2.5% improvement over the previous single-model state of the art.
The release targets a difficult area for AI systems, where subtle data definitions and complex business context can make natural-language database queries unreliable. Google said BIRD measures execution-verified accuracy, meaning Gemini-SQL2’s SQL not only appears correct but also runs successfully.
From the sources (3 posts)
@googleresearch🚀 Introducing Gemini-SQL2, our breakthrough text-to-SQL capability powered by Gemini 3.1 Pro! We've achieved state-of-the-art results on the highly competitive BIRD benchmark, translating natural language into execution-ready SQL queries. 🧵
@googleresearch📊 Data subtlety & complex business contexts make generating accurate SQL from natural language notoriously hard. Per the BIRD benchmark, which measures execution-verified accuracy, GeminiSQL-2’s SQL doesn't just look right, it also runs
@mirrokniProud of the team behind Gemini-SQL2 and collabotors from Cloud Research. +2.5% improvement over previous SOTA for singe model! 👇 @yanbang_wang, @qitianwu_, Sami Abu-El-Haija, Mohammadreza pourreza, @michael_galkin, @hemmatihadi, Hailong L