Experimental benchmark
Try Qwen in your browser
Compare a one-letter answer with a short JSON reply on your own device.
This runs Qwen 0.5B locally. Both formats use all 24 layers. It tests the cost of writing an answer, not the early-stop classifier.
Starting downloads about 800 MB from Hugging Face and the runtime from jsDelivr. The browser may keep the model in its cache. Allow a few minutes and keep this tab open; use Stop to end the worker. The three test messages and results stay in this browser.
Ready. Nothing is downloaded until you start.
Outputs and timings
A = card delivery; B = stolen card; C = incorrect cash withdrawal.
| Request | Format | Output | Expected | Check | Tokens | Time |
|---|
What this comparison measures
Each format gets one warm-up, then two rounds of three short authored requests. Their order alternates. One-letter replies are capped at one generated token; JSON at sixteen. Incorrect labels and invalid output formats count as failures. The prompts share the task but differ in their output instruction.
Times include tokenization and generation, excluding download, model loading and warm-up. This is a small timing demonstration, not an accuracy benchmark. The trained classifier and early-exit CPU tests are separate.
The model uses four-bit weights and one CPU worker through WebAssembly. Memory limits or browser restrictions can prevent it from running. No result is uploaded.