Hi Devs!
I’ve been testing the gemini-3.1-flash-lite model via the AI Studio interface. While it performs exceptionally well there, I’m noticing a significant drop in quality when applying the exact same configurations in the Vertex AI console.
The core task is straightforward: extracting specific numbers from text snippets. In AI Studio, this works perfectly across a large test suite. However, the Vertex AI responses contain illogical inconsistencies.
Has anyone else experienced this discrepancy? I tested a few other models and found the same issue—results from generativelanguage consistently outperform aiplatform