START_HERE.md

Start Here (3 Minutes)

Use this path if this is your first time with WebLLM Bench.

1) Open the app

• Live: https://ar5en1c.github.io/webllm-bench/
• Local: python3 -m http.server 8765 then open http://localhost:8765/bench.html

2) Pick the right preset

In Bench tab, use Preset:

• Desktop Chrome/Edge (recommended default)
• Safari Stable (lower defaults, still reproducible)
• Mobile Experimental (small windows for stability)

3) Run your first reproducible benchmark

• Model: Qwen2.5-1.5B-Instruct-q4f16_1-MLC
• Profile: 1024/128/10
• Enable: Force full max tokens
• Click: Run Benchmark

4) Export evidence

• Click Export JSON
• Keep browser version + screenshot + JSON together

5) Long-context check (8k workflow)

• Expand Add Custom Model and click ⚡ Add Qwen2.5 8k Preset (Hosted)
• Compare custom 8k model vs 4k baseline using prompt >4096 tokens
• 8k should run; 4k should overflow

References:

• docs/VALIDATION_PROTOCOL_8K.md
• docs/DEVICE_MODEL_GUIDE_2026-03-30.md
• docs/HUGGINGFACE_ONE_CLICK_8K.md