# Run a private browser LLM benchmark

Canonical page: https://browserllm.dev/run/

Choose a supported model artifact and exact encoding, runtime version, workload, token budgets, sampling settings, warmups, request schedule, seed and safety limits. WebGPU and sufficient browser/GPU memory are required; requirements vary by artifact. Model weights download from their pinned upstream hosts. Inference runs locally in a browser worker. Prompts and outputs are not sent to a model API.

The primary comparison loads one model once and runs one warmup followed by 60 measured requests in three windows of 20. Each request starts immediately after the previous response completes. There are 90-second loaded-idle gaps after windows one and two. After the third window: 30 seconds loaded idle, unload, then 15 seconds observation. Conversation history resets between requests. The windows share one session and are not independent trials. Older reports can use different schedules: read each report's saved protocol before comparing.

Browser capability checks describe exposed APIs, not guaranteed compatibility. A public browser run does not include privileged native CPU/GPU/power collectors. Finished reports remain in the current browser's private local storage, with View report and Download JSON actions. They never automatically update the curated leaderboard.