Mensura measures local AI inference performance directly on your Android device.
Run a repeatable benchmark against a pinned on-device language model and inspect the complete
test configuration instead of relying on an opaque score. Mensura reports decode speed, time to
first token, prefill speed, model-load time, generation latency, peak process memory, and thermal
status.
TRANSPARENT RESULTS
Every saved result includes the device, Android version, model, runtime, backend, workload,
validation outcome, and run configuration needed to interpret the measurement.
WHAT IT MEASURES
• Decode tokens per second
• Time to first token and prefill speed
• Median and p95 generation latency
• Model-load time and peak process memory
• Thermal state and deterministic workload validation
REPRODUCIBLE BY DESIGN
• Explicit CPU configuration with no silent backend fallback
• Configurable warmup and measured-run counts
• Live benchmark progress and streamed model output
• Local, app-private benchmark history
• Shareable benchmark result images
LOCAL-FIRST
The first benchmark can download the pinned model after showing its name and download size and
receiving your approval. Once verified, the model remains in app-private storage and later
benchmarks can run offline. A local model-file import is also available as an offline fallback.
Mensura has no account, ads, analytics, telemetry, or cloud inference. It is intended for Android
developers, ML/runtime engineers, device reviewers, and QA or device-lab teams.
Results describe the tested device × model × runtime × backend × configuration. They are not
universal device scores.
Benchmark local AI inference on Android with transparent, reproducible metrics.