Start Here: Device + Model Guidance

• The hosted Qwen2.5-1.5B ctx8192 preset is included directly in the model dropdowns as a built-in option.
• Use Chrome/Edge for benchmark comparability; Safari is supported but slower in our published runs.
• Default stable choice: Qwen2.5-1.5B-Instruct (q4f16_1) for balanced quality/speed on laptops.
• Need faster decode and can spend more memory: test Qwen3-1.7B.
• For prompts beyond 4k context, use the custom ctx8192 build and follow the 8k validation protocol.
• Mobile is experimental: start with prompt ≤512 and max output ≤128.

🔧 Add Custom Model

Model Weights URL (optional if local files selected)

Local Model Files (optional)

Drag/drop local model files or a folder (`mlc-chat-config.json`, `tokenizer.json`, `tensor-cache.json`, `params_shard_*.bin`).

Choose Files Choose Folder Clear

No local model files selected.

Local files are session-only (browser security). Re-add after reload.

Model Lib URL (optional if local wasm selected)

Drag and drop a local .wasm file here (no link needed for model lib).

Choose .wasm Clear

No local .wasm selected.

Model ID (Unique Name)

Context Window

VRAM Required (MB)

➕ Add to Registry 🗑 Clear Custom Models