ludion-bench measure browser inference on your device

This device

probing…

Full capability table

probing…

Device label (optional — joins the public dataset)

Live

Results

Median per engine × model × prompt × cache

Advanced — manual plan (operator workflows)

Plan

Engines WebLLM Transformers.js wllama

Models & order Qwen2.5-1.5B → Llama-3.2-1B (desktop) Llama-3.2-1B → Qwen2.5-1.5B (iPhone) Qwen2.5-1.5B only Llama-3.2-1B only Qwen2.5-0.5B only (floor)

Auto-continue next session after reload

“Run cold” wipes Cache Storage / IndexedDB / OPFS first. The HTTP disk cache cannot be cleared from JS — for a guaranteed cold run, also clear site data in browser settings.

Export

Operator notes (thermal feel, throttling suspicion, …)