$ bestmodel.run — workbenchpool snapshot 2026-09-188,230 runs · 686 models · 184 rigsevery number declares its basis
What do you want to run?
Intent, machine, quantization and context are four separate decisions, so they get four separate controls. Nothing is fused and nothing is estimated — a combination the pool has never tested says so.
01 what you want to run
Ordered by how much the community has tested it.
03 quantization
04 context floorFilters to cells community-tested at least this far.
255.9tok/s · Qwen3.8-27B-W4A16-AutoRound
The verdict above is the pool answering, live from 8,230 community runs on real hardware. Frozen snapshot 2026-09-18 — basis printed beside every number, ranking provisional. Read the pool yourself or start from hardware.
Every number declares its basis
$ basis --explain
measured median of ≥3 single-stream runs on this exact cell
reported 1–2 runs · real, but thin
claim (unvalidated) best community claim for the same intent on another rig · orientation only, always labeled
extrapolated scaled by memory bandwidth · never shown as measured
no data yet nobody has run it · we say so
The loop that feeds the engine
Nothing on this diagram is planned — every node is a component that ships in the repo today, and the engine is only as honest as the loop that feeds it.