Tags
In a pilot, Apollo 2 humanoids sorted car parts on Mercedes-Benz lines at a 96% success rate using Gemini Robotics 2.5, which is rolling out to robotics partners only. At 96%, would you put the robot on the line or keep a person checking every tray?
Gemini 4 Carbon moved past its cyber-defenders-only release and is live in Agent Studio, per a Threads post. Google’s own pitch gives it a 1M output token limit. Have you ever needed anywhere near 1M output tokens in a single run?
1,048,576 input tokens is the ceiling for Meta’s Muse Spark 1.3 on Google Cloud, but the preview skips batch predictions and Provisioned Throughput. Would you put a preview model with no batch mode into production, or wait?
In Google’s own tests, moving Qwen3’s KV cache took about 230 milliseconds with TPU Sync instead of several seconds, though Google warns the library’s APIs still change fast. Would you put a library with fast-changing APIs in production, or wait for stable?
