Actual WASM inference compared with CPU PyTorch. No recorded responses are used for generation.
Ready