LOCAL BROWSER BENCHMARK

Run a browser benchmark

Choose a model, weight format and request schedule.

Browser capabilities

Browser and device information, checked locally. No capability data is uploaded.

Checking browser capabilities…

Configure a run

Your results stay in this browser. View or download the report when the run ends; nothing is added to the public leaderboard.

Find a model

Model

46 matching models · 19 with browser packages · 33 catalog candidates at or below 8B parameters

Weight encoding

SmolLM2 135MWebLLM 0.2.82 · WebGPU · dedicated worker

Weight files: 269 MB
Estimated GPU memory: 359.7 MB

Previously benchmarked artifact. Memory is a catalog estimate, not a reservation or a measurement.

Checking WebGPU…Capability detection does not guarantee that the model fits or runs successfully.

Workload

Run schedule

20 requests → 90 s idle → 20 → 90 s idle → 20

One request at a time, back-to-back within each window. Chat resets before every request; the model stays loaded between windows.

Output token limit

Context window

TemperatureTop-pSeed

Reasoning mode

The context budget includes the prompt and output. Output may end before the token limit. Compare runs with the same settings.

Weights download only when you run. Cached files may be reused.

Run monitor

IDLE

Ready to run

Select a model and schedule, then start the benchmark.

0:00 total0 / 60 measured requests

GENERATED OUTPUTAnswer accuracy is not tested

Model output will appear here during the run.

What gets measured

¹ First-output latency: time from request start to the first non-empty text chunk in the worker. ² Generation speed: decoding tokens per second, reported by WebLLM. The summary above shows the first measured request; the full report includes all requests.

Actual token counts, finish reasons, warmups and timing boundaries are saved with the report. Basic output checks do not establish answer accuracy.

Loading…

What gets measured

Actual token counts, finish reasons, warmups and timing boundaries are saved with the report. Basic output checks do not establish answer accuracy.

Run a browser benchmark

Configure a run

Generation settings

Warmup, resource sampling & limits

Prompts & model files

Run monitor

Ready to run

What gets measured

Run a browser benchmark

Configure a run

Generation settings

Warmup, resource sampling & limits

Prompts & model files

Run monitor

Ready to run

What gets measured