Topic
Running models locally on devices at the point of data generation rather than round-tripping to cloud data centres.
Cactus released Whistle, an open-source speech recognizer for seven languages that runs in 16.9 MB on a CPU with no GPU. Its speed lead over Whisper base was measured on one Apple chip, so teams building for wearables and microcontrollers need their own numbers.
Reality
Liquid AI's open-weight d1-3B answers classification questions in a single forward pass, taking 16 ms on a Jetson AGX Thor and 50 ms on an Orin Nano. That puts routing and intent tagging on edge boards without a large model, on figures Liquid measured itself.
Perspective Coverage
3 publishersReality
MLCommons' MLPerf Inference v6.1 adds tests that time a whole RAG pipeline and a 1,007-turn coding-agent replay. Those scores help a buyer size hardware only where the buyer's own pipeline spends its time the way the four-model reference does.
Reality
Forum Markets formed a 51%-owned venture with Edge Node to own and operate inference compute across 12 to 13MW of sites that already have power and permits. It plans to sell to large, creditworthy companies on multi-year terms, and for those buyers the date compute switches on matters more than proximity.
Reality
Ukraine says it recovered an Nvidia Jetson Orin NX 16GB from an S-71M cruise missile. The load-bearing fact is not edge inference; it is that a 2023 module reached a market Nvidia left in 2022.
Reality
Cactus Compute released Whistle, an open 16.9MB speech-to-text model that runs on a CPU in the same Needle engine as its tool-calling language models. The accuracy and speed figures are the company's own, so they need rerunning on real devices and audio.
Reality
NVIDIA's CUDA buffer backend lets a standard ROS 2 message field carry GPU memory, and the same host, the same device, the same Linux user and a supported RMW all have to line up before the copies actually disappear.
Reality
New York Times reporting on the wreckage of a July 2026 strike on a Zaporizhzhia gas station says operators programmed the drone toward the area and the onboard AI chose the object it hit. Three civilians died.
Reality
NVIDIA's submission runs the same Qwen3.6-27B as the llama.cpp reference on the same Jetson board and finishes 6.4x sooner. Most of the gap comes from prompt tokens the runtime never has to prefill.
Reality
Nemotron 3.5 Lightning touches a tenth of its weights per token, which is what makes an agent loop plausible next to the sensors, but memory still has to hold all thirty billion and NVIDIA's own loop still escalates off-device.
Reality
The published spec quantifies pins, rails and access sizes but never a data rate, which tells you the decisions device makers face for the next several years are about granularity and board cost rather than headline bandwidth.
Reality
The first hardware disclosure from the leading robotaxi operator turns an autonomy pitch into a supply-chain document, with 20x compute growth in eight years and redundancy it calls non-negotiable.
Perspective Coverage
3 publishersReality
A Pixel 10 Pro wired into a Formula E GEN4's telemetry bus ran a local agent and fed the driver advice in about a second. The architecture is the story, not the hillclimb time.
Reality