01 — the question
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 — if the pool has never tested a combination, it 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
02 — the answer
Your machine already has an answer.
You just asked it. What came back is not a spec-sheet estimate — it is what 8,230 community runs on real hardware actually recorded, with the basis printed beside every number.
03 — the pool
Measured beats reported.
A cell becomes measured at three runs. Below that it stays visibly reported, and a combination nobody has tested renders as no data yet rather than as a plausible-looking estimate.
8,230
runs in the pool
686
models indexed
184
reference rigs
Frozen snapshot 2026-09-18 — not live throughput.
04 — the ladder
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