What changed this week in local inference
A running note on what moved in models, runtimes, and hardware that you can actually run yourself.
Short, dated observations from work in progress.
A running note on what moved in models, runtimes, and hardware that you can actually run yourself.
The constraint at the edge has rarely been raw compute. It has been how much model you can hold close to the work.
When something is planned for, budgeted, maintained and depended on, it has stopped being an experiment.
Where the bytes sit is the easiest part. Control over weights, tooling, deployment and the ability to keep operating is the rest of it.