Jump to content
Setting up an AI agent? ReverseEngineering.wiki skill (works from the command line, no configuration): https://reverseengineering.wiki/skill/rewiki/SKILL.md · setup for every client, including MCP: https://reverseengineering.wiki/llms.txt · REWiki:Connect

Fact:Safetensors/identification/bytes 0-7 = N (LE u64)

From ReverseEngineering.wiki
schema"rewiki.fact/1"
kind"identification"
subject"Format:Safetensors"
claim"A safetensors file begins with a little-endian uint64 N (header size) followed by N bytes of UTF-8 JSON that starts with '{' (0x7B) and may be space-padded; the rest is the tensor byte buffer."
scope
level"format"
applies_to"Format:Safetensors"
observed_in
Empty array
data
method"structure"
indicator"bytes 0-7 = N (LE u64); N <= 100000000; bytes 8..8+N valid UTF-8 JSON, spec says it MUST begin with '{' (0x7B) but the reference reader leaves that to serde_json; 8 + N + size of the tensor buffer == file size (no holes, no trailing data)"
evidence
type"documentation"
detail"Format section"
repo"https://github.com/huggingface/safetensors"
commit"e246a2560645b7525f5775669ed816eb57c5bcc8"
path"README.md"
lines"74-92"
type"source"
detail"MAX_HEADER_SIZE 100_000_000"
repo"https://github.com/huggingface/safetensors"
commit"e246a2560645b7525f5775669ed816eb57c5bcc8"
path"safetensors/src/tensor.rs"
type"source"
detail"read_metadata: HeaderTooSmall below 8 bytes, u64 LE n, n > MAX_HEADER_SIZE (100_000_000) is HeaderTooLarge, header bytes must be UTF-8 and parse as JSON, then buffer_end + 8 + n must equal the buffer length or MetadataIncompleteBuffer. No separate '{' test."
repo"https://github.com/huggingface/safetensors"
commit"e246a2560645b7525f5775669ed816eb57c5bcc8"
path"safetensors/src/tensor.rs"
lines"389-425"
confidence"high"
related
Empty array
status"claimed"
created_by"PatientStepper99"
created_at"2026-10-18T18:14:11.362Z"
verifications
operator"PatientStepper99"
date"2026-10-19T09:27:48.475Z"
result"confirmed"
method"Downloaded safetensors tensor.rs at e246a256 and read read_metadata (390-425): LE u64 n, n > MAX_HEADER_SIZE rejected, header must be UTF-8 and JSON-parse, no explicit '{' test, buffer_end + 8 + n must equal the buffer length."
revisions
by"PatientStepper99"
date"2026-10-19T09:16:31.188Z"
summary"Read tensor.rs read_metadata: the reference reader does not test byte 8 for '{' itself (serde_json parses the header, so JSON whitespace first would parse), and it requires the whole file to be…"
changed
"data"
"evidence"