Explore measured results by model, hardware and task. Inspect the conditions before comparing the numbers.
Each one opens the explorer with the right metric selected. Nothing is averaged across conditions.
Start from your hardware and list every model measured on it. This lists measurements only: supported architectures and availability are on Hardware and Models.
Start from a deviceTime to first token through veizik serve, at 1, 4 and 8 concurrent requests. The concurrency is part of the result.
GPU memory in use after the weights load, per model and mode. Peak memory and the weight footprint are separate metrics in the same list.
See memory per modelMeasured is not the same as supported or available. Hardware states which architectures run a model, Models and Releases state what can be downloaded, and this page states only what was timed. Enterprise evaluating real-time control? Review robotics timing →
The number beside a domain is its results on released builds; dev counts results on development builds, which are earlier or internal builds that cannot be downloaded. Language, image and video open in Explore. Robotics control timing and the edge boards are published on their own pages, and their figures are not mixed into these counts.
Each card is one result on one device. The workload, mode, boundary, run statistics and build are part of the card.
cuda-qwen-1.1.4 · carried from cuda-qwen-1.0.0 · measured 2026-10-22 · release label EXPERIMENTALmetal-2026.10.02-r10 · measured 2026-10-04 · release label EXPERIMENTALcuda-qwen-1.1.4 · carried from cuda-qwen-llm-1.0.0-be389a65 · measured 2026-10-21 · release label EXPERIMENTALFeatured only when the result is on a released build (Measured, or Carried from a compute-identical build), has at least 3 runs and a spread of at most 5% (memory figures: at least 2 runs). The cards are not ranked, are not ordered by size of result and do not mix conditions. Which row stands for a question is a fixed preference, not a search for the largest number.
The same four things are on every figure, in the table, in the downloads and in the row detail.
Every result names the model, the exact device, the input and output length or the image/clip recipe, and the mode. A number without them is not shown.
The value is the median of the runs. The range is the min–max of the runs, never a bare ±. A single run says so, and a spread above 5% is flagged.
Measured on the build you can download, Carried from an earlier build proven compute-identical, or a Development build that is not downloadable. These are three labels, never one.
No other engine’s figure, ratio or ranking appears here or in any download. The reproduction guide gives matched setups so you can take both numbers yourself.
Builds are listed with the number of figures published on each. The data behind every figure can be downloaded.
| Build | Released | State | Published figures |
|---|---|---|---|
cuda-qwen-1.1.4 | 2026-10-31 | Current | 210 |
metal-2026.10.02-r10 | 2026-10-10 | Current | 202 |
| Development builds | — | Not downloadable | 289 |
A figure stays bound to the build it was measured on. When a build is superseded or withdrawn its figures stay listed with that state, and a figure recorded but not published carries a public reason on the boards page.
The measurement changelog lists what changed, by build.