The lab
Why we build local models
We want people to be able to run and adapt the models we build. That puts hardware requirements and operating control in the research brief.
We want a developer to be able to test a model on their own work, choose its version and decide where its inputs go. Local inference gives them that option when the model fits their hardware.
reflex-1 runs on a laptop CPU. Its main adaptation run trained 9.04M parameters inside a 421M-parameter serving model. That makes it possible to study decision tasks without updating the entire network. The report gives the accuracy and timing results, including where the checkpoint failed.
Samsara asks a different question: how much manipulation performance can we get from a 217M-parameter policy? Its 99.6% LIBERO Object result is promising. The training report also records the eight H200s used for the run. Model size alone would understate that cost.
We want the software around these models to leave decisions inspectable. An operator should be able to see which model ran, what it selected and which action followed. They should be able to test a replacement before using it.