Hugging Face is shipping 207 Apache-2.0 WebGPU kernels as individually versioned Hub repos, plus an npm loader (`@huggingface/kernels@preview`) that resolves a kernel by repo ID and contract version, derives output shapes from a manifest, and picks among variants (vectorized, broadcast, scalar) per call. Each repo carries its own correctness cases, benchmark cases, and WGSL Jinja templates, so the artifact is inspectable without reading shader code. The headline claim: 2.57x geometric-mean speedup over ORT WebGPU on an Apple M4, across 809 of 1,756 original test cases — they kept only the ones with matching outputs and reliable timings. Add 3.52x, Softmax 2.11x, LayerNorm 2.22x, MatMul a more honest 1.14x. The outliers are spicy: a bilinear Einsum at 10,000x (0.136ms vs 1,396ms) and a CumSum at 301x. Those are general implementations falling off cliffs, which is precisely the argument for specialized, per-shape kernels. Two caveats worth holding: they timed GPU work only, excluding setup and readback, which is where browser inference usually bleeds; and 947 excluded cases is a lot of fog around that 2.57x. Fleet, the crowdsourced in-browser benchmarking suite, is the real long game — device variance is where WebGPU lives or dies, and no M4 test lab covers it. The 300-baud kid in me notes we've come full circle: hardware-specific hand-tuned kernels, distributed through a central repo, downloaded on demand. We called that "shareware." Dario is presumably drafting 15,000 words on how crowdsourced GPU telemetry could end civilization, then shipping his own on Thursday.