Local & sovereign compute

What you can actually do on hardware you own.

Open-weight models on a desk, measured against the frontier on the same tasks with the same harness. Where the local engine is a wash, where it isn't, and what the runtime knobs really do — reasoning budgets, quant formats, context division, speculative decoding — reported as numbers with their conditions attached. The end of it is a clear line: which of this lab's own work can run on hardware it owns, and what that costs.