Google AI Updates: October 20, 2026

1. Google Releases the LiteRT Tensor API for Building Edge Models Op by Op in C++

Google. Google released the LiteRT Tensor API, a C++ authoring layer for building model graphs op by op, including LLM composite ops such as scaled dot-product attention, RMSNorm, and KV-cache updates. Graphs can load weights directly from Hugging Face .safetensors checkpoints without conversion scripts, export to .tflite for LiteRT on CPU, GPU, or NPU, or run directly on CPU through XNNPACK without a model file, and share accelerator buffers so pre- and post-processing avoid CPU copies. In Google’s SAM 2 demo, WebGPU throughput rose from under 10 FPS to over 30 FPS at 512x512, and on a MacBook Pro the same C++ code ran about 1.5x faster than Apple MLX in fp16. Python authoring, direct ML Drift GPU backends, and automatic full-graph quantization are listed as planned work, and the post does not state a beta or stable status. Source

2. Google Details ADK’s Non-Blocking Tools for Gemini Live Voice Agents

Google. A Google for Developers guide shows how the Agent Development Kit (ADK) keeps voice agents talking during slow tool calls, using streaming tools that yield progress updates and FunctionTool with response_scheduling modes (WHEN_IDLE, INTERRUPT, SILENT). In a sales-report benchmark with gemini-3.8-live, time to the agent’s first words fell from 14.4 seconds with a blocking tool to 1.2 seconds with a streaming tool, while the backend query itself still took 13 seconds. The patterns require google-adk 2.10.0 or later and run_live bidirectional sessions, and a built-in stop_streaming function cancels running tools. The authors also recommend a “call once” docstring directive plus an idempotency key to stop models from repeating tool calls. Source