Google AI Updates: October 21, 2026

1. Google Cloud Launches Gemini for Government Services in Preview

Google. Google Cloud released Gemini for government services in preview, bringing its Gemini agent to federal, state, and local agencies as part of a new suite of industry-specific solutions. The offering has five components: skills written by domain experts that encode statute, regulation, and agency guidance; connectors to case management, ERP, HR, and federal data sources that query systems in place without migrating data; pre-built partner agents for tasks such as FOIA discovery and batch redaction; a governed control plane with VPC Service Controls, CMEK, and Assured Workloads; and a partner ecosystem of integrators and public sector ISVs. Outputs are grounded in source documents with citations, and Google says a person approves each consequential step in partner agent workflows. Source

2. Google Launches Native PyTorch vLLM and SGLang Engines on TPU, Built on TorchTPU

Google. As part of a set of PyTorch announcements, Google launched native PyTorch backends for vLLM (developed with Inferact) and SGLang (developed with RadixArk) running on TorchTPU, which makes the TPU a native torch.device("tpu") target. Both engines keep each framework’s scheduler, KV cache manager, and API server, and add TPU-specific Pallas kernels, KV cache management with TPU Sync, and disaggregated serving. Separately, Google and Meta published a TorchTitan fork for TPU with training recipes for dense, MoE, and diffusion models such as Llama 3 405B, Qwen 3 235B MoE, and FLUX.1, so PyTorch teams can train on Cloud TPUs without rewriting code in JAX. Performance claims on Ironwood TPU are Google’s own. Source Source

3. Google Research Completes the Open Coral NPU Architecture and Ships Coral Compiler v1.0

Google. Google Research reached three milestones for Coral NPU, its open RISC-V based neural processing unit for edge devices: the complete 32-bit architecture including a new Matrix Execution Unit configurable for up to 1 TOPS while targeting 10 mW average inference power, and the open-source Coral Compiler v1.0 based on IREE/MLIR. The compiler takes models from JAX, PyTorch, and TensorFlow directly to Coral NPU hardware without custom assembly. Developers can prototype on the Synaptics Astra SR200 MCU series and Coralboard dev kits. Source

4. Claude Opus 5.5 and Sonnet 5.5 Added to Google Antigravity

Google. Google Cloud made Anthropic’s Claude Opus 5.5 and Claude Sonnet 5.5 available in Google Antigravity for Google Cloud customers, selectable from the model picker once an administrator enables them. The models work across Antigravity 2.0, the Antigravity CLI, and the Antigravity IDE extensions for VS Code, JetBrains IntelliJ, and Visual Studio, under unified enterprise billing alongside Gemini 3.8 Flash and Gemini 4 Argon. Google’s post also describes multi-model patterns, such as using one model family to implement and another to review, and assigning Flash-tier subagents for high-volume tasks via .agents/agents/*.md files. Source

5. Google AI Edge Model Explorer Adds Support for Custom Compiler Graphs

Google. Google AI Edge’s Model Explorer now supports custom compiler intermediate representations and open-source adapters for TOSA, VGF, and PTE formats used in on-device machine learning. New features include hierarchical graph visualization, split-pane views, and per-node trace overlays to catch export regressions, verify hardware delegation, and find latency bottlenecks locally. Source