🧠 Load Qwen 2.5 1B RLCD in your browser
Le modèle tourne localement via WebGPU / WASM — no server, no API key, no data sent anywhere.
First load downloads ~600 MB of quantized weights (cached after that).
📝 Context / Input Text
📋 JSON Schema
enum + boolean fields only
Temp: 0.2
🐢 Autoregressive Baseline
token-by-token
Waiting...
⚡ Parallel Constrained (RLCD)
1 forward pass
Waiting...
🔬 How does this work?
🐢 Autoregressive (Standard)
Standard LLMs decode JSON one token at a time. A 4-field schema might require 80–500 sequential forward passes through the model, each blocking on the previous.
[Context] → "{" → "\n" → " " → "risk" → ":" → " " → "HIGH" → ...
(requires 80–500 sequential passes)
⚡ Parallel Constrained (RLCD)
RLCD exploits that enum/boolean fields have a bounded candidate set. It broadcasts a single KV-cache state to evaluate ALL fields simultaneously — one pass regardless of schema size.
┌→ [Field 1: "risk_level"] → pick best
[Context KV] ─┤→ [Field 2: "urgent"] → pick best
└→ [Field N: "action"] → pick best
(1 forward pass, O(1) tokens)