Fine-tuned, quantized 1–10B models brought useful assistants to phones and PCs without a full cloud round trip.
Phi, Gemma, quantized Llama-3-8B, and similar models showed on-device utility for summary, keyboards, and local agents.
Silicon vendors (Apple, Qualcomm, Intel, NVIDIA) optimized kernels. Privacy and latency drove adoption.
The hybrid stack—edge + cloud—consolidated as the default product architecture.
MLCommons announced on 16 September 2026 MLPerf Inference v6.1 results with a record 30 organizations, new End-to-End RAG and Edge Agentic Inference tests, and up to 5.7× year-on-year gains in some scenarios—a different angle from the Vera Rubin NVL72 piece.
Labs showed models that click, type, and navigate UIs like humans, extending automation without APIs.
Enterprises deployed agents with real tools—CRM, code, browser—under tracing, permissions, and continuous eval.