Mentatcurated
Open Models

Open Models

Open weights, local inference, and running real models on your own machine.

State of the world · updated June 2026

Right now: open weights trail the closed frontier by months, not years, and increasingly run on a single GPU. Quantisation and small, capable models make local-first genuinely practical.

Watch: the next open release closing the gap, on-device tooling maturing, and the live tension between democratisation and proliferation.

Start here · the primer

Every few months the open frontier closes the gap to the best closed model, then someone runs it on a laptop. This is the democratisation track — and the proliferation worry — at the same time.