Wattever

Local models. Real numbers.
detecting… LM Studio
THIS MACHINE

System

Hardware inventory and live readings from this computer.

THIS MACHINECPUDetecting...GPUDetecting...RAMUnreportedVRAMUnreportedDetails
Storage probe settings optional

Use a folder on the model drive. Paths stay private.

Existing folders only; up to eight additional drives.

Live hardware

Waiting for a sample.

YOUR LOCAL BENCHMARK

See what your setup can do.

Choose your models and workload. Keep one clean result to share.

Model configuration Shared with results

Actual engine settings, reported by you. Leave unknown values blank.

Hardware compute claims

Optional documented peaks or independently measured throughput. Quantization bits do not establish compute precision. Use TFLOP/s for floating point and TOPS for integer arithmetic. Each claim needs a source and a dense/sparse distinction.

1Models

Choose from your downloaded models.

2Workload

Uses reasoning_effort. Unsupported values fail visibly; choose Provider default to omit it.
Select each size to test.
Off is the standard benchmark. Compare both runs every test twice.
Advanced settingsOptional
Leave blank to omit.
MODEL CAPABILITIES

Understand every result.

Explore strengths, failure modes and consistency. Select a chart to inspect its evidence.

New capability evaluation

Uses your endpoint and model configuration from Run.

Run settings
Thinking + final answer
0 = no wall-clock deadline

These settings belong to this capability evaluation. Generation budget includes reasoning and the final answer. Requested reasoning does not prove the engine disabled thinking. A server context limit can still stop generation; inspect output limits and coverage before comparing scores.

YOUR EVALUATION WORKSPACE

Start with a result.

Run a suite or select a saved evaluation. Pass rates, category profiles, repeat consistency and question-level evidence appear here.

YOUR BENCHMARK LIBRARY

See what your models can handle.

Compare total throughput, each request's experience and the cost of serving more users.

IMPORT AND EXPORT

Share

Exchange benchmark and capability results.

Benchmark results

Select benchmarks in Results to export their measurements as CSV, or open a benchmark to download its JSON.

For contributions, use the individual download buttons above to get ready-to-upload {id}.json files. Bundles must be split into separate files before upload.

Share a benchmark

Download the benchmark JSON, then upload it to results/runs/ in a fork and open a pull request. Once merged, it will appear in the shared library.

Contribution guide

Capability evaluations

Select evaluations in Capabilities, then use the individual download buttons below to get ready-to-upload {id}.json files. You can also export a bundle or import JSON with the desktop helper. Bundles must be split into separate files before upload.

Export bundle

Add or rerun tests

Choose models, techniques and an action for this benchmark. Existing measurements are kept unless you choose to rerun them.

Merge benchmarks

Benchmarks that ran on one machine become one benchmark. Each engine keeps its own models, sessions and settings, so engines compare side by side.