Private data room
The supplied data licence permits internal use by the authorized account holder. Enter the separate AI Bench access code to load benchmark data.
The API key stays on the server. Raw data is never available before authorization and is not indexed.
Price / form factor
Choose a catalog view, then select one or more GPU/devices from that class.
PASS targets ≤8 seconds after warm-up. FAIL also covers closed/unrecognized models and unsupported accelerator runtimes. Catalog .
GPU / device
A model is shown when it fits at least one selected device. All selected or none selected both mean the whole class.
Local-fit estimate: warm, batch 1, ≤256 input tokens, ≤80 output tokens, ≤4K context, text only, reasoning off, Q4/native 4-bit, no CPU offload for discrete GPUs. Hosted Artificial Analysis speed is not used for this filter.
Models in view—
Index version—
Peak intelligence—
Peak output speed— tok/s
Signal field
Intelligence ↑ · output price → · dot size = speed
selected
Comparison rail
0 of 4 models selected · Amber marks selection leaders.
Swipe sideways for pricing and performance →
| Select | ||||||||||
|---|---|---|---|---|---|---|---|---|---|---|
| Loading… | ||||||||||
—