AI developer tools
We provide an end-to-end AI toolchain for developers to access and maximize our AI infrastructure.
Harness the power of TPUs
Find the resources you need to build, train, and serve AI on TPUs. Accelerate your entire life cycle, from pre-training to production-ready serving with vLLM, JAX, and PyTorch.
Resilient training, post-training, and inference at scale
Mitigate hardware failure with elastic training using MaxText
Scale agentic reinforcement learning training with Tunix
Build production AI on TPUs with JAX
Optimizing Qwen 3.5-397B MoE on Ironwood (TPU7x)
This systems engineering playbook details how to maximize inference throughput using first-principles roofline analysis, custom Pallas kernel fusions, and SparseCore-accelerated routing.
Google AI Edge
Deploy AI across mobile, web, and embedded applications
Start now with task APIs, run LLMs on-device, or deploy custom models from Google AI Edge.
Integrate task APIs into your app
Add prebuilt AI features like background blur, face mesh, and object detection.
Run LLMs on-device
Run the same LLM across Android, iOS, web, and embedded devices.
Deploy a custom model across platforms
Convert and deploy hardware-accelerated models from PyTorch, JAX, TensorFlow, and Keras.
Enable agentic assistance via local models completely offline
Build with the same agentic capabilities that power Google Antigravity, with initial support for Gemma 4 26B A4B using Google AI Edge’s LiteRT.
Manage complex AI operations and data flow
Develop and deploy Graph Neural Network (GNN) models
Use Graph Flow, a Python toolbox to train, evaluate, and deploy GNN models and other relational data ML models.
Develop and deploy state-of-the-art ML algorithms
Use TensorFlow as your end-to-end open-source ecosystem of tools and libraries to easily build, train, and deploy AI/ML applications.
Migrate faster with AI-assisted code
See how Google pioneered 6x faster migration from TensorFlow to JAX by deploying specialized, multi-agent AI systems.
Leverage Kubernetes, the foundational open standard for AI
Advanced relationship with specialized hardware
Enhanced scale and reliability with GKE and core capabilities
Optimized performance for efficient inference
Simplified usability and enhancing observability
Run and scale AI agents at high density
Use Agent Substrate, an open-source, secure-by-default execution runtime purpose-built for agents. It delivers 10x higher sandbox density and sub-500ms suspend/resume activations with native zero-trust isolation.
Use Kubernetes AI infrastructure building blocks
Security & sandboxing
Use gVisor, our open-source application kernel providing secure container isolation (the foundation of GKE Sandbox).
Job scheduling & batch queuing
Orchestrate batch, ML training, and multi-tenant compute resources with Kueue, our Kubernetes-native queue manager.
AI post-training & fine-tuning
Run RL/SFT loops on Kubernetes clusters with OpenRL, our self-hosted post-training API from GKE Labs.
Agentic operations & automation
Deploy autonomous AI agents to help you manage, audit, and remediate Kubernetes fleets with Kube-agents, an autonomous agentic harness.
llm-d is a Cloud Native Computing Foundation Sandbox project
We’re proud to be a founding contributor to llm-d, the fastest path to state-of-the-art LLM inference on any accelerator, uniting around a clear, industry-defining vision: any model, any accelerator, any cloud.