September 2025 Summaries

Tunix is a newly introduced, open-source, JAX-native library designed for post-training alignment of large language models (LLMs) that simplifies the transition from pre-trained models to production-ready systems. It offers a comprehensive toolkit for aligning models at scale, specifically optimized for performance on TPUs, with features including supervised fine-tuning, preference tuning, knowledge distillation, and advanced reinforcement learning methods like PPO, GRPO, and GSPO. Tunix's "white-box" design allows developers full control over the training process, enabling easy customization without dealing with complex abstractions. It seamlessly integrates with the JAX ecosystem, providing modular and user-friendly APIs for common post-training workflows, while its initial release supports various algorithms and techniques for model alignment and compression. Developed in collaboration with academic and industry partners, Tunix is praised for its flexibility, ease of use, and ability to address real-world challenges in model alignment and agentic AI, with the community encouraged to contribute and collaborate on its development.

EmbeddingGemma is an advanced text embedding model derived from the Gemma model family, specifically adapted from a pretrained Gemma 3 model. It utilizes numerical representations called embeddings to interpret text meaning and context, making it effective for tasks like search and retrieval. The model transforms the original decoder-only architecture into an encoder-decoder format, optimizing its ability to generate expressive embeddings by employing a combination of three loss functions: Noise-Contrastive Estimation, Global Orthogonal Regularizer, and Geometric Embedding Distillation. These techniques enable EmbeddingGemma to produce robust, high-quality embeddings suitable for various applications, with flexibility in dimensionality to balance performance and efficiency. Additionally, the development process includes a multi-stage training approach, incorporating methods like Quantization-Aware Training and Model Souping, which enhance the model’s quality and versatility for semantic technology advancements in areas such as Retrieval-Augmented Generation and hyper-personalization.

Grounding with Google Maps in Vertex AI is now generally available, providing developers with tools to create generative AI applications that incorporate real-world, up-to-date information from Google Maps. This feature enhances the reliability of AI outputs by leveraging verifiable geospatial data and user reviews, facilitating timely and contextual responses to user queries such as business hours or local recommendations. It is particularly beneficial across various industries, including travel, real estate, and social media, by enabling enhanced location-based insights and personalized user experiences. For instance, real estate companies can tailor property summaries to client profiles, while travel services can offer detailed trip recommendations. The integration of Google Maps with AI enables more comprehensive and dynamic AI responses, such as combining data from Google Search and Maps to answer complex questions. Currently available in the United States and India, this capability empowers developers to experiment and implement these features using the Gemini API in Vertex AI.

The GKE Inference Gateway, an extension of the Google Kubernetes Engine Gateway, optimizes the deployment and management of generative AI workloads by enhancing routing, load balancing, and observability. It supports dynamic Low-Rank Adaptation (LoRA) model serving and integrates AI safety checks with Google Cloud Model Armor. The gateway also allows for model-aware routing and the prioritization of latency-sensitive requests. To address enterprise demands for secure and optimized AI workloads, the GKE Inference Gateway integrates with Apigee's API management platform through the GCPTrafficExtension resource, providing comprehensive governance and monetization of Agentic APIs. Apigee's robust features, such as policy enforcement, API lifecycle management, and advanced analytics, enable organizations to effectively manage and monetize their AI services, with plans to further enhance AI policy governance, including model security and semantic caching.

The integration of the Agent Development Kit (ADK) with AG-UI offers a streamlined way for developers to create interactive AI applications by connecting robust backend capabilities with a flexible frontend protocol. ADK provides an open-source toolkit for building AI agents that can reason, plan, and execute tasks, while AG-UI acts as an open protocol and UI layer that facilitates rich, real-time interaction between these agents and users. This combination allows developers to build sophisticated AI agents with minimal effort, featuring shared state and human-in-the-loop capabilities, which enhance user interaction by allowing agents to generate and render UI components directly in chat and react to user actions. The integration is supported by CopilotKit, an open-source library of React components, enabling developers to quickly deploy polished chat interfaces and other UI elements. By alleviating the complexity of connecting backend logic with user interfaces, this integration focuses on delivering powerful, intelligent, and collaborative AI applications efficiently.

Updated versions of Gemini 2.5 Flash and 2.5 Flash-Lite are now available on Google AI Studio and Vertex AI, improving quality and efficiency by reducing output tokens, with the Flash-Lite version achieving a 50% reduction and Flash a 24% reduction. The new Gemini 2.5 Flash-Lite is enhanced with better instruction following, reduced verbosity, and stronger multimodal and translation capabilities, making it more suitable for high-throughput applications. Meanwhile, the Gemini 2.5 Flash model shows significant improvements in tool use and cost-efficiency, achieving higher quality outputs and gaining positive feedback from early testers, including an increased performance on agentic tasks. These updates are part of an ongoing effort to iterate and refine the Gemini models, with the introduction of a -latest alias for easy access to the newest versions, while also allowing users to provide feedback and build production-ready experiences. The release strategy emphasizes testing and feedback to shape future stable versions, with a 2-week notice period for any updates or deprecations to ensure users can adapt accordingly.

Gemini Robotics-ER 1.5, now available in preview via Google AI Studio and the Gemini API, is a pioneering model designed to enhance robotics with advanced embodied reasoning capabilities. This model excels in visual and spatial understanding, task planning, and progress estimation, making it adept at handling complex tasks that require contextual information and multiple steps, such as sorting objects into recycling bins based on local guidelines. It is optimized for rapid spatial reasoning, generating precise 2D points, and orchestrating advanced agentic behaviors through spatial and temporal reasoning. Users can control the latency versus accuracy trade-off, allowing the model to think longer for complex tasks or respond quickly for simpler ones. Additionally, Gemini Robotics-ER 1.5 includes improved semantic safety filters and physical constraint awareness, ensuring safer operation within defined parameters. As a high-level reasoning engine for robots, it integrates with various tools and APIs to execute sophisticated tasks, demonstrating significant performance on both academic and internal benchmarks.

Running large language models (LLMs) directly on devices like Chrome, Chromebook Plus, and the Pixel Watch is made possible through LiteRT-LM, a framework designed for efficient and high-performance on-device inference. This approach offers the advantages of offline availability and cost-effectiveness, eliminating per-API-call costs and making LLMs practical for frequent tasks such as text summarization and proofreading. LiteRT-LM addresses the challenges of deploying gigabyte-scale models across various hardware by utilizing a modular, open-source design that supports multiple platforms and accelerators, including CPU, GPU, and NPU. The framework's architecture, consisting of an Engine and Session system, allows shared resources to be managed efficiently while enabling customization through lightweight adapters. This system enhances flexibility and scalability, adapting to different device constraints, from powerful smartphones to resource-limited wearables like the Pixel Watch, where a minimal pipeline can be constructed to optimize binary size and memory usage. The framework also integrates with Google's broader AI Edge stack, supporting developers in building custom LLM-powered applications and scaling them across diverse platforms, highlighting its utility in products like Chrome and the Pixel Watch's Smart Replies feature.

Google Colab has introduced several new features aimed at enhancing the educational experience by making notebook-based learning more robust and reproducible. A key addition is the Runtime Version Selector, which allows users to pin notebooks to specific Colab runtime versions, ensuring consistent execution and preventing compatibility issues caused by software updates. This feature is particularly beneficial for researchers, educators, and developers who require stable, verifiable results. Additionally, Colab has improved its slideshow mode, enabling presentations to start from any cell and allowing pairing of collapsible headings with code cells for seamless demonstrations. New URL link features facilitate easy sharing and copying of notebooks, further supporting classroom and collaborative workflows. These updates are part of Colab's ongoing commitment to enhancing user experience by providing a stable, shareable, and reproducible platform for machine learning, data science, and education.

The public release of the Data Commons Model Context Protocol (MCP) Server represents a significant advancement in offering AI developers, data scientists, and organizations quick and actionable access to Data Commons' extensive public datasets. This server aims to mitigate Large Language Model hallucinations by anchoring them in real-world statistical data and streamlining the development of data-rich applications. It enables AI agents to consume Data Commons natively, thus accelerating the development of trustable, data-driven applications, which are capable of handling a wide range of queries from exploratory to generative. A notable application of the MCP Server is the ONE Data Agent, developed in collaboration with Google's Data Commons and the ONE Campaign, which facilitates the rapid and intuitive search of vast health financing datasets. This tool aids in global health efforts by providing accessible insights into health financing, thereby enhancing advocacy, reporting, and policy-making. The MCP Server is designed for seamless integration into agent development workflows within the Google Cloud Platform, offering minimal onboarding friction, and is supported by resources such as the Agent Development Kit, Gemini CLI, and sample agents in Google Colab.

Gemini CLI now integrates with FastMCP, a leading Python library for creating Model Context Protocol (MCP) servers, simplifying the development and deployment of custom MCP tools and prompts. This integration allows users to install local STDIO transport MCP servers using the command fastmcp install gemini-cli server.py, which automatically handles configuration and dependencies while connecting directly to Gemini CLI. FastMCP's user-friendly Pythonic design, featuring decorators and type hints, facilitates the creation of tools and prompts, enhancing Gemini CLI's capabilities by allowing it to perform actions and access additional context. Users can easily incorporate these prompts as slash commands, streamlining their terminal interactions. The integration encourages developers to explore and create innovative MCP servers, with the option to contribute improvements or new features via GitHub for both open-source projects.

The 0.2.0 release of Google's Agent Development Kit (ADK) for Java introduces integration with the LangChain4j LLM framework, expanding the range of Large Language Models (LLMs) available for developers to build AI agents. This integration allows access to models from various third-party providers and local models via tools like Ollama or Docker Model Runner. Developers can specify models using the LlmAgent builder, which supports both string names and instances of classes extending the BaseLlm abstract class, creating a bridge between ADK and LangChain4j. The release also enhances agent development with improved tooling, asynchronous support, loop control, and advanced logic and memory features, such as chained callbacks and memory management services. These updates aim to make ADK a more flexible framework for creating robust AI agents, and the release encourages new users to explore the possibilities with comprehensive documentation and community support.

Genkit Go 1.0 is the first stable release of Google's open-source AI development framework designed for the Go ecosystem, offering a robust platform for building full-stack AI-powered applications. This release introduces the genkit init:ai-tools command to enhance AI-assisted development workflows, providing a unified interface for multiple AI model providers and streamlined APIs for diverse functionalities such as multimodal content, tool calling, and retrieval-augmented generation. Genkit Go allows developers to create and deploy production-ready AI applications efficiently, ensuring type-safe AI flows with Go structs and JSON schema validation. It includes rich local development tools, a standalone CLI binary, and a Developer UI, facilitating fast and intuitive AI application development. The framework supports integration with popular AI coding assistants, enhancing the development experience by simplifying debugging, testing, and code generation. With Genkit Go, developers can leverage AI to build scalable and reliable applications, benefitting from Go's inherent speed and safety.

The Gemini Batch API has been enhanced to support the new Gemini Embedding model and offers compatibility with the OpenAI SDK, allowing developers to submit and process batches more efficiently. This update builds on the initial launch of the Gemini Batch API, which enables asynchronous processing at reduced rates, making it ideal for high-volume and latency-tolerant use cases. The Gemini Embedding Model, already in use for numerous production deployments, can now be leveraged with the Batch API at higher rate limits and half the price, facilitating cost-sensitive and asynchronous applications. Additionally, developers can integrate the OpenAI SDK compatibility by modifying a few lines of code, enabling easy transition to using the Gemini Batch API. The enhancements aim to optimize costs and expand the utility of the Gemini API, with further updates anticipated.

At Google I/O, Google introduced Gemma 3n with text, image, and now audio capabilities, expanding the Google AI Edge stack to support audio through the MediaPipe LLM Inference API for Android and Web. This development enables features like speech-to-text and speech-to-translated-text without requiring an internet connection, and it is available in the Google AI Edge Gallery app, which has reached 500,000 APK downloads. The app, now in open beta on the Google Play Store, serves as an interactive platform for developers to experiment with on-device AI models, offering practical examples and source code available on GitHub. Future plans include releasing the app for iOS, showcasing more generative AI tasks, transitioning to the LiteRT-LM runtime for flexibility, and collaborating with the community to optimize and publish models to Hugging Face LiteRT. The initiative is supported by a diverse team of contributors and aims to inspire and assist developers in building AI-powered applications.

JAX, a framework initially known for AI model development, is gaining traction in various scientific domains due to its capability to efficiently handle complex transformations and derivatives, particularly in fields like physics-informed machine learning. Researchers Zekun Shi and Min Lin from the National University of Singapore and Sea AI Lab have successfully leveraged JAX's Taylor mode automatic differentiation to address computational challenges in solving high-order Partial Differential Equations (PDEs), which are difficult to manage with traditional frameworks optimized for backpropagation. Their innovative Stochastic Taylor Derivative Estimator (STDE) allows for efficient computation of high-order derivatives without the exponential cost typically associated with such tasks, achieving significant speed and memory improvements. This breakthrough, which earned a Best Paper Award at NeurIPS 2024, underscores JAX's potential as a versatile and powerful tool for scientific computing, enabling researchers to solve previously intractable problems and advancing the role of differentiable programming in scientific discovery. The researchers' work highlights the growing importance of JAX in scientific research, beyond its established applications in deep learning, and calls for continued community engagement to further enhance its capabilities for scientific advancements.

The A2A protocol offers a structured framework for agent-to-agent communication, emphasizing flexibility through extensions that allow customization beyond its core protocol. These extensions enable developers to add domain-specific functionalities and requirements, facilitating specialized features on A2A servers and clients. The protocol's open ecosystem, identifiable by unique URIs, is community-driven, promoting adaptability in multi-agent systems. Notable extensions include a traceability feature that enhances collaboration by tracking agent interactions and Twilio's latency-aware extension for optimizing voice agent performance. Additionally, Identity Machines utilizes zero-trust handshakes for secure task delegation, while Ethereum's ERC-8004 aims to establish a trust layer for AI agents, fostering secure, cross-organizational communication.

Veo has launched significant updates including support for vertical format videos, 1080p HD output, and reduced pricing for Veo 3 and Veo 3 Fast to enhance accessibility and quality for video generation. These updates make Veo 3 and Veo 3 Fast stable and ready for production in the Gemini API, with price reductions making high-quality video creation more affordable. The platform now supports aspect ratios of 9:16 for mobile and social media use, and allows for higher quality outputs with 1080p resolution. Developers can leverage these capabilities in various applications, such as Invisible Studio and Saga, which utilize Veo 3 for efficient video generation workflows in AI-driven content creation, while Mosaic utilizes agentic workflows to edit raw footage. Veo 3’s advanced features, including superior prompt adherence and fluid motion, have quickly become essential for productivity in short-form content creation. These updates also include a digital SynthID watermark on all generated videos, and resources like documentation and starter apps are available to facilitate the integration of these new features.

Google's new embedding model, EmbeddingGemma, featuring 308 million parameters, is designed for efficiency and versatility, making it ideal for both on-device and cloud applications, particularly in semantic search and Retrieval Augmented Generation (RAG). This model enables streamlined knowledge ingestion pipelines when integrated with Google Cloud's Dataflow and vector databases like AlloyDB, facilitating the conversion of unstructured data into embeddings and their subsequent storage in vector databases. The model's open nature allows for secure, large-scale data processing entirely within Dataflow, eliminating the need for external services and improving operational efficiency. EmbeddingGemma is fine-tunable for specific data needs and ranks highly in multilingual text-only models on the MTEB leaderboard. The integration of EmbeddingGemma into a Dataflow pipeline offers improved efficiency, scalability, and simplicity by processing data locally on Dataflow workers, thus avoiding remote procedure calls and reducing resource footprint. Dataflow's 'MLTransform' simplifies the pipeline creation process, enabling the generation of embeddings and their storage in vector databases like AlloyDB with minimal code, enhancing the capability to develop advanced AI applications.

EmbeddingGemma is an innovative open embedding model designed for efficient on-device AI applications, boasting state-of-the-art performance for its size with 308 million parameters. This model is specifically tailored for use in environments that require privacy and offline capabilities, such as mobile devices, by enabling tasks like Retrieval Augmented Generation (RAG) and semantic search without needing an internet connection. With its foundation in the Gemma 3 architecture, EmbeddingGemma supports over 100 languages and is capable of operating with less than 200MB of RAM due to quantization techniques. It offers flexible output dimensions and a 2K token context window, making it suitable for a variety of devices and use cases. EmbeddingGemma integrates seamlessly with popular tools such as sentence-transformers and transformers.js, facilitating easy adoption for developers. By leveraging Matryoshka Representation Learning, it allows for multiple embedding sizes, optimizing both quality and performance. The model underscores its utility in providing high-quality text embeddings crucial for on-device applications, ensuring secure and efficient processing of sensitive data while enabling new capabilities like offline personal file searches and customized chatbots.