Daisuke Majima (MLBoy)

Idol (speridlabs) on a Galaxy S26: one photo becomes a 3D object you can turn around, sides and back included. The encoders, the 5-step turbo denoiser and the Gaussian decoder run on the phone's GPU with LiteRT. Graphs, runner and Android sample: huggingface.co/litert-communi…

LiquidAI's d1-3B in Edge on a Windows laptop via WebNN.
Type a state and your questions; it returns calibrated answers in one forward pass each.
About 16 ms a question on the RTX once warm.
The page builds the prompt, tokenizes and reads answers itself.
github.com/john-rocky/d1-…

U²-Net Portrait, one ONNX file, three engines on one Core Ultra 9 386H laptop chip via OpenVINO.
Per sketch, median of 10: Intel GPU 27 ms, NPU 39 ms, CPU 309 ms.
Mean difference from the CPU sketch: 0.06 (GPU), 0.05 (NPU) of 255 gray levels.
More models:
github.com/john-rocky/win…

Qwen3-0.6B on an RTX 5080 Laptop: TensorRT Edge-LLM built from source on Windows vs llama.cpp CUDA, same text.
This run: decode 381 vs 322 tok/s, fp16 weights on both.
Benchmark medians, 369 vs 322:
github.com/john-rocky/ten…
Independent work, not sponsored or endorsed by NVIDIA.

MNN-Pocket-0.3B on a Galaxy S26's NPU, from LiteRT-LM's command line over adb.

Same model, same question, two files. Start-up: 0.27 s precompiled for the phone's chip, 26.66 s compiled on the phone.
Same answer. The wait plays at 4×; times are real.

huggingface.co/litert-communi…