421M-parameter Laya model runs in a browser tab, download up to 478 MB
A developer tested the open 421M-parameter decision model Laya in a browser via ONNX Runtime Web. First load takes 26–35 seconds (327–478 MB), while a three-question call runs in 358 ms on WebGPU and about 930 ms on CPU.
- 8-bit build weighs 478 MB, 4-bit 327 MB; load times 35.4 s and 25.9 s
- Three questions take 358 ms on WebGPU and 923–931 ms on CPU
- 4-bit quantization shifts answers: 10 of 12 matches versus 11 of 12 for 8-bit
- In a background Chrome tab the same call takes 6.2–13.2 s instead of 0.92 s
Read next
AI