PCC beats xz-6 on Whisper weights. 1,178,964 bytes. Verified.

Whisper Tiny FP32. 151,061,672 bytes in. PCC wrote 67,103,604. Ratio 0.444. Every byte verified by decode plus SHA-256.

The baselines on the same file: xz-6 68,282,568. zstd-19 74,556,043. gzip-9 86,795,913. zstd-3 88,176,968. All 16 benchmark cells (4 weight files × 4 compressors) decode-verified. PCC beats the strongest baseline by 1.73%.

Exact selection, no shortcuts. Seven modes — LZ, BWT, column-BWT, transposed-BWT, LZM2, per-block mixing, xz — each encoded the file for real, and the smallest actual byte count won. Mode 8 (per-block exact selection) took this file.

The honest costs: encode took 3,809 s against xz-6's 258 s — about 15× slower. Decode 36 s vs 10 s. A win without costs is marketing; this is a ratio-vs-speed tradeoff, stated plainly. Need fast, use zstd. Need small, PCC just took the crown on this file.

One file of four. GPT-2, SmolLM2, and Whisper Base baselines are done; their PCC runs are pending. No broad claims about neural weights in general. One verified win is one verified win.

Why it matters: model weights are the most-moved bytes in ML infrastructure — every checkpoint download, every deployment, every edge device. At that scale, 1.7% is real bandwidth and real money.

Full four-file benchmark and paper draft in progress, targeting the IEEE BigData 2026 Data & Model Efficiency workshop.