Paper Referee Report

What “Choose model” means

The “model” is the AI brain used to generate the referee feedback. It runs locally in your browser using WebGPU (no API key, no server upload). Bigger models usually give sharper, more “referee-like” critiques, but take longer to download and need more memory.

Recommended picks

• Best quality (recommended): Qwen2.5-7B-Instruct-q4f16_1-MLC
• Fastest / easiest to load: Llama-3.2-3B-Instruct-q4f16_1-MLC
• Even stronger critique (if available): choose an 8B Instruct model (heavier but better)

Quick tips

• Prefer Instruct models (they follow directions better)
• q4f16 is usually the best speed/quality balance
• If you see context window errors, the PDF text is too long for the model settings