# Official Qwen FP8 PLE, reused for Swift1.5

These four files and the 2-byte BF16 scale were copied from the local Qwen
FP8 sidecar after complete Swift/Qwen BF16 PLE identity was proven. The
weights, scale and manifest are unchanged from the Qwen release.

Original source: `Qwen/Qwen3.8-Flash-Next-FP8@236dfdf285828023ca3bcd3f37366c58a3469b13`.
Storage: `FP8_E4M3FN`, 128 logical parts, 160-byte rows, 4,096-byte alignment,
shared BF16 multiplicative scale `0x3951`.

Decode each E4M3FN code to BF16, multiply by the BF16 scale, and round to BF16.
The runtime keeps compressed pages in a bounded cache.

`verify-extraction.json` and `verify-independent.json` are the original Qwen
extraction receipts. `hashes-independent.json` records the fresh copied-file
verification for this package. Run `sha256sum -c SHA256SUMS` from this directory.

See the [root card](../README.md) and [reproduction](../reproduction/README.md).