Qwen2.5-0.5B-Instruct, running on WebGPU. No PyTorch, no transformers, no
tokenizer library: the safetensors reader, the bfloat16 conversion, the BPE tokenizer, the
KV cache, the attention kernels and the quantization are all in
the repository.
Generating plausible text proves nothing; the reference
implementations already do that. Every claim below is checked against an independent
implementation or measured on named hardware, and the ones that went against me are
here too.