data: Llama 4 Scout BF16 shard 5 → bgz17 (18.2 GB → 7.7 MB, 4735×) Streamed from HuggingFace via HTTP range reader. Zero disk for source. MoE expert FFN: 15,420× compression. Shared expert: 964×. Attention: 2,162×. Full model estimate: ~215 GB BF16 → ~40 MB bgz7. https://claude.ai/code/session_01Y69Vnw751w75iVSBRws7o7#49