Nvidia found one training step changes just 1% of a model.
Nvidia researchers published a method on Tuesday that copies only the weights a training step actually changed. Syncing a trillion-parameter model between two cloud regions fell from 87 minutes to 150 seconds. Nvidia says the receiver ends up with bit-identical weights, and the code sits in its open NeMo library.
Why it mattersTraining agents means shipping a fresh model to the machines that test it, over and over. Cutting an hour of waiting out of each round makes better agents cheaper to build.
Read in the app — free, in 9 languages
Related stories
South Korea's president says AI helped hackers into four banks.2026-10-07Meta and Sierra proposed one sign-in rule for all shopping agents.
2026-10-07OpenAI posted 722 maths papers from one prompt to one agent.
2026-10-07OpenAI's new endpoint picks from a fixed list in 150 milliseconds.
2026-10-07The Erdős problems site stopped taking proofs after an AI flood.
2026-10-07