Benchmarks

How we score a recipe

Compatibility on this site means a recipe passed the gates we print — not a VRAM spreadsheet. These are our definitions. We do not copy another registry’s wording or hidden cutoffs.

Load

The process attaches to the discrete card (rocminfo / --list-devices agrees). We do not publish a time-limit SLA.

Chat ends cleanly

Generate ends. The server does not hang mid-token. This is not a quality score.

Reasoning split

A distinct thought block and a final answer both appear. Not measured on our seeded recipes yet.

Tool call

The server emits a tool call and consumes the result. Not measured on our seeded recipes yet.

Long-context recall at a stated fill

The run used the context we printed (262144 for the daily driver). We do not publish a planted-token recall score. We do not publish a planted-token percentage until we run that test.

Decode recorded

tok/s at the stated context, sourced. No invented floor. There is no invented “must beat N tok/s” cutoff on this page. A slow honest number beats a missing one.

Blank card

Every seeded recipe starts with the unmeasured boxes honest.

Loads on the named GPU Not measured

The process attaches to the discrete card (rocminfo / --list-devices agrees). We do not publish a time-limit SLA.
A reply finishes Not measured

Generate ends. The server does not hang mid-token. This is not a quality score.
Thinking channel (when the model has one) Not measured

A distinct thought block and a final answer both appear. Not measured on our seeded recipes yet.
A tool call is used Not measured

The server emits a tool call and consumes the result. Not measured on our seeded recipes yet.
Stated context actually ran Not measured

The run used the context we printed (262144 for the daily driver). We do not publish a planted-token recall score.
Decode recorded Not measured

tok/s at the stated context, sourced. No invented floor.

Back to benchmarks · amdai-bench/run@1 cards