Nanbeige engine laboratory
Explicit runtime comparison. All inference stays on your GPU.
Engine
Custom WGSL
ONNX Runtime Web
Model
Compact calibrated · 2.35 GB
Q8 embedding calibrated · 2.60 GB
Original Q4 control · 2.60 GB
Subgroups
Shader f16
Context
512
1024
2048
Prefill chunk
64
16
8
1
Packed download
Kernel tuning
Baseline
32-token tile
GEMV R4
Fused norm R4
Fused norm R8
Explain why the sky is blue in two sentences.
Load
Run 64 tokens
Benchmark suite × 3
Validate 6 prompts + cache
Stop
Export JSON
Ready.