Continuing to Bring You Our Latest Models: Improved Gemini 2.5 Flash and Flash-Lite Release Date: September 25, 2025 Authors: Shrestha Basu Mallick, Sid Lall, Zach Gleicher, Kate Olszewska (Google DeepMind Product Team) --- Overview Google DeepMind announces updated versions of the Gemini 2.5 Flash and 2.5 Flash-Lite models, available on Google AI Studio and Vertex AI. These releases aim to deliver higher quality outputs with improved efficiency and reduced costs. --- Key Improvements and Highlights Updated Gemini 2.5 Flash-Lite This latest Flash-Lite model focuses on three main areas: Better instruction following: Significantly improved ability to follow complex user instructions and system prompts. Reduced verbosity: Produces more concise answers, leading to lower token usage and reduced latency—important for high-throughput applications. Stronger multimodal and translation capabilities: Enhanced audio transcription accuracy, improved image understanding, and better translation quality. Model for testing: gemini-2.5-flash-lite-preview-09-2025 --- Updated Gemini 2.5 Flash Improvements target: Better agentic tool use: Enhanced performance in complex, multi-step, agentic tasks. Notably, a 5% performance gain on the SWE-Bench Verified benchmark (48.9% → 54%). More efficient: Achieves higher output quality while using fewer tokens, reducing latency and costs. Early testers praise the update for its speed, intelligence, and cost-effectiveness, enabling scalable AI applications such as Manus, an autonomous AI agent. Model for testing: gemini-2.5-flash-preview-09-2025 --- Testing and Development Google encourages developers to test these preview models to experience the latest innovations and provide feedback. These preview releases are not yet stable versions but help shape future stable models. New Model Aliases To simplify access, Google introduces a -latest alias for each model family, pointing to the most recent version to avoid frequent code changes: gemini-flash-latest gemini-flash-lite-latest Developers will receive two weeks' notice before any updates or deprecations to the -latest alias. For production stability, continue using the stable version strings: gemini-2.5-flash gemini-2.5-flash-lite --- Visuals Summary Quality and speed comparison charts show improvements for Gemini 2.5 Flash and Flash-Lite preview models over stable releases. Token efficiency visuals demonstrate a 50% reduction in output tokens (costs) for Flash-Lite and 24% for Flash compared to prior models. --- Additional Resources & Related Posts Building the Next Generation of Physical Agents with Gemini Robotics-ER 1.5 Google Colab Adds More Back to School Improvements! Introducing the Data Commons Model Context Protocol (MCP) Server --- Google continues to push the boundaries of AI with Gemini. The latest releases represent ongoing progress, inviting developers to build, test, and contribute as the models mature. Happy building!