[rwkeyes]

Problem and result

"Why is the 28.8 GB lookup table in a section about speed?" — it isn't a speed reason, and asking the question exposed
a real error next to it.

• The table is a capacity trade: it is read from the SSD a few rows at a time to avoid holding 28.8 GB in memory.
  Both versions do it. As a "why it is fast" bullet it was neither a speed reason nor a fork difference, so it is out
  of that list, which is now the three decisions that actually are about speed.
• It is, though, 28.8 GB of the download, and What you need said "roughly 40 GB" for Qwen3.8-Flash-Next. The
  model guides give 66–84 GB for the Q2_0/IQ3 packs — the pack plus this table. The disk row now says that, and a
  paragraph under it explains what the table is and links the derivation.

Related discussion

none — documentation only.

Approach

1. Why it is fast: three bullets (expert split with CPU/GPU parallel work, prompt-chunk reads, MTP guess-then-check),
   followed by the paragraph that says all three are stock behaviour, names the fork's only drafting settings
   (spec=N, spec_window, neither used by the measurement) and points at the server-side comparison.
2. What you need: memory row 38–55 GB (37.6 GB for the smallest pack); disk row 66–84 GB for
   Qwen3.8-Flash-Next, 239–274 GB for DeepSeek V4.1-Flash.
3. Under the table: the 28.8 GB is the model's lookup table, the publisher's second file, read from the SSD a few rows
   at a time, present in every pack of the model and in both versions; the derivation lives in the glossary.
4. Glossary keeps the full provenance and adds that the fork's only PLE source change is ple_parity, the table's
   self-test, which now skips instead of failing when the model file is absent.

Validation

Passed:

• Disk and memory figures taken from docs/MAC.md ("the original Q2_0 and IQ3 packs download roughly 66 to 84 GB",
  Q2_0 37.6 GB, IQ3_S 54.8 GB) and docs/MODELS.md#pick-by-memory (37.6 / 39.2 / 47.0 / 54.8 GB of RAM + VRAM).
• The table's size still closes to the byte from the GGUF header (160 × 320,001,536 × 18/32 = 28,800,138,240 B).
• Links, anchors, rendering: no broken relative link, the glossary anchor resolves, GitHub's Markdown API parses
  7 tables and 2 images.
• The old understatement is gone: grep for "roughly 40 GB" returns nothing.

Failed:

none observed.

Not run:

• No measurement or build; text and figures only.

Limits

• The 66–84 GB is for the original Q2_0/IQ3 packs; Unsloth packs are much larger and the page already sends the
  reader to docs/MODELS.md for the pack-specific number.