■ space · a field guide in five spreads

■ the story in one breath

GGUF is a self-describing box that makes quantization the default — small files, the llama.cpp shelf. safetensors is a table of contents taped to raw weights — full precision by default, the Python shelf. The runtime you already run picks your box; the disk picks your size. Speed is a property of runtimes and hardware — the bytes refuse to testify.

01THE WEIGH-IN

Same weights in every box. Rank them by size in your head before the bars fill — then look.

The gap is not waste. The 4-bit box forgot on purpose — each weight stored at a quarter of the bits, close enough for most work and a third of the disk. Keep that thought; spread three gets its own chapter. For now — pick a box up:

HOLDING: GGUF · Q4_K_M — 397 MB — the disk's favourite

02INSIDE THE CRATE

Every format is a layout. Here is each box, laid end to end — flip between them and read like a runtime does.

magic:bytes 0–3 are literally "GGUF" — a runtime checks this first
version:u32 little-endian — 3 today
n_tensors · n_kv:u64 counts — the file describes itself end to end
one file:header and weights together — carry it anywhere, it opens
header_len:u64 little-endian — how many bytes of JSON follow
json header:dtype, shape, byte offsets — a table of contents
then:raw weights, nothing else — safe to mmap straight off disk
born for:the Python stack — from_pretrained is the native tongue

segments not to scale — in both files the weights dwarf the paperwork · layouts from the public format specs

03THE FORGETTING

The runt's secret: each weight stored at a fraction of the bits. Drag the width and watch precision leave — the bytes follow.

4-bit · ≈ 397 MB
Q4 · 397 MB — shipped BF16 · 1.20 GB — shipped
on a shipped rung — the record's own bytes

At 2-bit the world snaps to 0, 1, −1 — a model that knows three things. At 16 it remembers every decimal it was born with. Between them sits your disk budget. What the bytes cannot say is what quality lost on the way down — that wall stands, and every honest table marks it.

04THE TWO SHELVES

The format decides nothing except this: which shelf opens. Say what answers on your machine — the room will answer back.

THE LLAMA.CPP SHELF

the GGUF boxes open here

Quantized single files, small disks, no Python in sight — the appliance shelf.

THE PYTHON SHELF

safetensors opens here

from_pretrained is the native path — raw tensors, no conversion step to maintain.

YOUR ROUTE: Ollama → the llama.cpp shelf → the GGUF crates

the hedge the record admits: vLLM reads GGUF through its plugin — GPU, with caveats — and Ollama, the GGUF appliance, still ships an MLX runner. even the shelves trade.

05THE VERDICT

You carried the GGUF · Q4_K_M crate the whole way. The disk's favourite — and with Ollama answering on your machine, it was never a contest.
GGUF makes quantization the default; safetensors makes full precision the default. That default is most of the decision — nobody ever switched format for speed. They switched runtime for a feature, and the format came along as the ticket. — the only verdict the bytes allow
RULE 01The runtime you already run picks the format.Check what answers on your machine before any table does.
RULE 02The disk picks the quant.A small box takes the 397 MB file and stops reading forums about it.
■ honest stamp I've read the record, not run it. Every number on this page is a shipped file's own bytes — 396,705,472 · 1,198,182,848 · 1,503,300,328 — pulled 2026-10-14, re-checked 2026-10-18, one curl -sIL … | grep -i content-length away from your own terminal. No bench rig. No tokens-per-second cosplay.

end of the yard · nothing fetched · everything local