johnnyteutonic's picture
pk_sft_v1 F16 (97.5M parameters, 221 MB) with ember.cpp 9d993f48de: greedy F16 = F32, WebAssembly = native, WebGPU = WebAssembly 40/40 (experiments/release_pk_sft_v1_q8_2026-10-28)
8e107b9 verified