Google Research proudly unveils pivotal 2025 I/O advancements, showcasing cutting-edge research, significant contributions to Google's Gemini models, and transformative generative AI product developments.
Each year at Google I/O, we present Google's most sophisticated technologies, demonstrating their utility and ability to unlock novel user experiences. We empower developers and communities with tools for innovation. Many of these pioneering technologies originate from extensive research within Google Research, often in deep collaboration with other divisions, building upon successive breakthroughs in artificial intelligence and computer science. This year's I/O emphatically underscores the power of translating groundbreaking research into tangible reality. As Sundar aptly stated: “This progress signifies our entry into a new era for the AI platform shift, where decades of dedicated research are now materializing for individuals, businesses, and communities globally.”
Beyond Google Research's foundational contributions to Gemini and the generative AI products highlighted at I/O, we are thrilled to present our select favorites this year. These innovations stem from years of dedicated effort within Google Research to realize the transformative potential of the research magic cycle.
MedGemma and AMIE: Revolutionizing Healthcare Through Advanced AI
Since the groundbreaking introduction of Med-PaLM in 2022, followed by Med-PaLM2 and Med-Gemini, our research teams have relentlessly advanced AI capabilities to enhance healthcare accessibility and efficacy. At I/O, we unveiled MedGemma, Google's most powerful open-source model engineered for multimodal medical text and image comprehension. This model holds immense potential to accelerate the development of novel healthcare solutions.
MedGemma is meticulously built upon Gemma 3, serving as an essential foundation for developers creating advanced health applications, such as sophisticated radiology image analysis or concise clinical data summarization. Its compact architecture ensures efficient fine-tuning for specific requirements. Critically, when assessed against the MedQA benchmark, its baseline performance in clinical knowledge and reasoning tasks rivals that of significantly larger models. As an open-source model, MedGemma offers unparalleled flexibility, enabling deployment in developers' preferred environments, including Google Cloud Platform or local infrastructures. Both the MedGemma 4B and the 27B text-only models are now readily accessible via HuggingFace and Vertex Model Garden, as integral components of our comprehensive Health AI Developer Foundations (HAI-DEF).
MedGemma's foundational performance in clinical knowledge and reasoning mirrors that of considerably larger models.
MedGemma advances our innovative work, following the recent introduction of AMIE. Developed in partnership with Google DeepMind and showcased at I/O, AMIE stands as a sophisticated research AI agent designed for advanced medical diagnostic conversations. The latest multimodal iteration possesses the remarkable ability to intelligently interpret and reason with visual medical data, significantly assisting clinicians in achieving more precise diagnoses.
LearnLM: Establishing Gemini as the Premier Global Learning Model
For nearly two years, Google Research and cross-functional Google teams have collaborated extensively with educational specialists on LearnLM, a specialized family of fine-tuned models optimized for learning. At I/O, we announced that LearnLM will now be seamlessly integrated directly into Gemini 2.5, positioning it as the world's leading model for educational advancement. Our most recent technical report demonstrates that Gemini 2.5 Pro substantially outperforms competing models in applying learning science principles, making it the preferred instrument for educators. It features advanced STEM reasoning, sophisticated multimodal understanding, robust quizzing and assessment functionalities, and numerous other enhancements.
Furthermore, we have launched an innovative quiz creation feature within Gemini, meticulously designed and optimized for learning by our Research team. Students aged 18 and above can leverage Gemini to generate customized quizzes tailored to any subject matter, utilizing their class notes or course documents. Gemini provides constructive feedback and detailed explanations for both correct and incorrect answers.
Utilize our comprehensive LearnLM prompt guide to maximize Gemini's pedagogical value. For instance, instruct Gemini to assume the role of a biology instructor or to dynamically adjust text complexity for specific grade levels.
play silent looping videopause silent looping videounmute videomute videoIn addition to embedding pedagogical advancements within Google products, we are actively collaborating with partners to deploy the powerful capabilities of our LearnLM models in diverse educational settings. In conjunction with Kayma, we successfully piloted automated assessment for both concise and extensive content with thousands of students and educators in Ghanaian high schools, with strategic plans for scaling to a greater number of students and countries.
Gemma's Multilingualism and Efficiency: Ensuring Global Accessibility and Utility
Aligned with Google's core mission to universalize access to global information, we are intensifying research in multilinguality. This initiative guarantees LLMs produce reliable outputs across various languages, ensuring genuine utility for users worldwide. Two months ago, Google launched Gemma3; our research significantly propelled Gemma's expansion to over 140 languages, establishing it as the premier multilingual open-source model today. At I/O, we announced the integration of these advanced capabilities into Gemma3n, the latest addition to the Gemmaverse. This model is optimized to operate with as little as two gigabytes of RAM, making it ideal for on-device applications. Our persistent focus on efficiency ensures the Gemma3n model delivers reduced latency and superior energy consumption friendliness.
To empower developers in building and refining multilingual models, Google Research recently introduced ECLeKTic, an innovative benchmark meticulously designed to evaluate cross-lingual knowledge transfer within LLMs.
Efficient and Grounded Models: Fueling Advancements in Google Search AI Mode
As Large Language Models (LLMs) scale in size and user demand escalates, our capacity to enhance model efficiency while simultaneously elevating or maintaining their quality is paramount for democratizing access to these high-performance systems. Google Research has achieved significant efficiency breakthroughs that have become industry benchmarks, including our pioneering work on speculative decoding and cascades.
We have published extensive research on techniques for ensuring factual consistency techniques and rigorous evaluations. We have set a new standard for factuality and grounding through features like double-check and the FACTS Grounding leaderboard, developed collaboratively with Google DeepMind and Kaggle. Now, our research has been instrumental in enhancing the AI Mode within Google Search, significantly improving the user experience.
Announced at I/O, AI Mode represents Google's most advanced AI search capability, featuring sophisticated reasoning power. It is progressively rolling out to all U.S. users, empowering them to conduct more in-depth research through follow-up queries and direct links to pertinent websites. Our advancements in efficiency ensure that these models operate with enhanced reliability and deliver faster responses. Concurrently, our factuality research has refined web searching mechanisms within AI Mode, guaranteeing that delivered information is highly accurate, grounded in multiple authoritative sources, and accompanied by relevant links.
Multimodal Factuality: Enhancing Imagen4, Gemini 2.5, and AI Avatars in Vids
As multimodal content proliferates, our dedicated factuality team is pioneering advancements in multimodal factuality research to uphold stringent accuracy standards across all Google products. We have significantly elevated the quality of Imagen4 within the Gemini app, the latest image generation model unveiled at I/O, capable of producing visuals with remarkable lifelike detail. For AI avatars integrated into Vids—a novel feature enabling users to rapidly create video content featuring custom AI avatars—we played a crucial role in evaluating model quality and image caption accuracy. Additionally, we delivered substantial improvements to the video comprehension capabilities of Gemini 2.5 models, with a specific focus on high-motion understanding, thereby enhancing Gemini's proficiency in analyzing human motion across health and fitness domains.
Sparkify: Transforming Questions into Animated Visual Narratives
Our teams were instrumental in supporting the launch of the innovative new Labs experiment, Sparkify. This groundbreaking platform integrates the potent capabilities of Gemini, MusicLM, AudioLM, and Veo, empowering users to transform any query or concept into a concise and captivating animated video, rendered in their chosen design style. The project leverages the robustness and factuality of its underlying foundational models. Join the waitlist for an opportunity to experience Sparkify firsthand.
Watch the film
FireSat: Pioneering Early Detection of Smaller Wildfires
Continuing our long-standing commitment to mitigating the catastrophic impacts of wildfires, Google Research has forged a vital partnership with the Earth Fire Alliance, the Moore Foundation, and Muon Space to develop FireSat. FireSat is a groundbreaking satellite constellation engineered for expedited and more precise global wildfire detection. It leverages high-resolution, multispectral satellite imagery and AI to deliver near real-time intelligence for first responders and enable scientists and ML experts to meticulously study fire propagation dynamics. In March, we successfully deployed the inaugural satellite of this constellation, comprising over 50 satellites. This initiative significantly expands upon our established wildfire boundary tracking capabilities, which provide critical data through Search and Maps, and our synthetic Firebench dataset, released on the Google Cloud Platform to accelerate scientific research in this critical field.
play silent looping videopause silent looping videounmute videomute videoFireSat represents a paradigm shift in wildfire detection, utilizing high-resolution imagery for earlier identification.
Quantum AI: Unlocking Tangible Potential for Real-World Applications
During the Dialogues Stage session, Sinead Bovell, founder of WAYE, and Julian Kelly, Senior Director of our Quantum Hardware team, explored the profound promise of quantum computing and the remaining engineering and scientific hurdles. Julian highlighted the recent advancements from Google Research's Quantum AI team, including our state-of-the-art Willow chip and significant progress in critical areas such as quantum error correction. Complex computations previously intractable for classical computers can now be executed on quantum chips in mere minutes, paving the way for a multitude of transformative real-world applications. The potential to revolutionize fields like pharmaceutical development and energy efficiency is becoming increasingly concrete and accessible.
Furthermore, we developed an immersive Quantum AI game experience for I/O attendees: the Quantum Maze Runner. Participants raced against the clock to navigate the maze, subsequently witnessing how a quantum computer would efficiently solve it.
AI Co-Scientist: Accelerating the Pace of Scientific Discovery
Our revolutionary AI co-scientist, featured at I/O and meticulously developed in collaboration with Google DeepMind, operates as a sophisticated multi-agent system powered by Gemini. It excels at synthesizing complex information and executing advanced reasoning tasks. This tool is intentionally designed as a collaborative partner for scientists, empowering them to generate novel hypotheses, formulate research proposals, and substantially accelerate biomedical discoveries. It has already demonstrated significant potential in critical areas, including drug repurposing for acute myeloid leukemia and proposing hypotheses for novel treatment targets for liver fibrosis.
This initiative represents just one facet of our extensive efforts to expedite scientific research across the broader ecosystem. Our new Geospatial Reasoning program is dedicated to advancing public health, urban planning, integrated business strategy, climate science, and more. We are also making significant strides in neuroscience. Our recent publication on LICONN presents the first-ever methodology for comprehensively mapping neurons and their connections in brain tissue using standard light microscopes. Additionally, we released the Zebrafish Activity Prediction Benchmark (ZAPBench), enabling researchers to investigate the intricate relationship between structural neural connectivity and dynamic neural activity across an entire vertebrate brain for the first time. Our research also extends to genomics, aiding in the diagnosis of rare diseases. REGLE, an unsupervised deep learning model, assists researchers in discovering associations with genetic variants. Furthermore, we have open-sourced new DeepVariant models as part of a collaborative effort on Personalized Pangenome References, which demonstrably reduce analysis errors by 30% when examining genomes from diverse ancestral backgrounds.
Conclusion
The research advancements detailed herein represent a curated selection of the ongoing, high-impact work undertaken by Google Research teams who are consistently driving breakthroughs across diverse scientific domains and translating them into tangible applications. In this extraordinary golden age of research, the synergistic cycle between investigation and real-world deployment is accelerating in both speed and scope. I/O provided an unparalleled platform to showcase how this dynamic process generates profound positive impacts for individuals, businesses, scientific endeavors, and society at large.
Acknowledgements
We extend our sincere gratitude to the numerous teams and collaborators whose invaluable contributions have enriched this blog post and the groundbreaking work it represents.