v3.4.0

Export, Windows Links & Smarter Quantization Picking

Highlights

Copy what you already downloaded out as plain files for LM Studio, Ollama or
llama.cpp without re-downloading, a cache that works on Windows without
Developer Mode, download huge sharded models in batches, and pick EXL2/EXL3
bitrates that live on separate branches.

Export & Flat Files (#83, #91)

• hfdownloader export <repo> <dest> copies a cached repo out as plain
  files in its own layout — nothing is re-downloaded, and the export is
  independent of the cache. -F exports only some weights (tokenizers and
  configs always come along), -b picks a revision, and --mode hardlink
  shares the cache's disk space instead of copying (same drive only).
• Web UI: "Export as real files" on any cached repo, copied to
  <export-dir>/<owner>/<name>. Enabled with serve --export-dir <folder>;
  without it the API refuses, so it can't write anywhere else.

Windows: Hardlinks Instead of Missing Links

• Without Administrator/Developer Mode, Windows used to skip the cache's
  links entirely — files ended up only as blobs/<sha256>, invisible to
  Python, the HF CLI and you. hfdownloader now uses hardlinks there
  (no extra disk space). On drives without any links (FAT/exFAT) it copies,
  asking once first in an interactive terminal.
• New --link-mode auto|symlink|hardlink|copy (download and serve, config
  key link-mode).
• hfdownloader rebuild repairs caches written by older Windows builds
  (recreates the missing snapshot entries from the download manifest).

Downloading

• --shards 1-100 (or 1-50,120-185) downloads only those shards of
  split files, so a huge model can be fetched in batches (#90).
• Quantizations on branches (#94): EXL3/EXL2 repos that keep each
  bitrate on its own branch (turboderp/...-exl3: 4.00bpw,
  SC_6.00bpw_H6_V6; bartowski exl2: 4_25, 6_5) are listed with bits,
  head bits and size; analyze recommends one and -b <branch> downloads it
  (CLI, analyze -i and the web UI).
• Weight-format choice for every model type (safetensors, PyTorch, ONNX,
  TensorFlow, TFLite, Flax, OpenVINO, Rust) — e.g. whisper-tiny's
  recommended download is 148 MiB instead of 580 MiB of four formats.
• GGUF repos offer every mmproj precision; diffusers LoRAs are detected.

Web UI & API

• Pause button now appears on jobs that waited in the queue; download speed
  is shown; skipped files show as "skipped".
• Saving settings no longer wipes other config keys (cache-dir,
  stall-timeout, ...), validates values (bad values used to break every
  later job) and writes the config file privately (0600).
• A second download of the same repo with different filters is its own job;
  deleting a repo's cache while it downloads is refused (409).
• Analyze and plan honor your proxy, return 404/401 for missing or gated
  repos, and no longer cut off slow analyses of huge repos.
• Request bodies are capped at 1 MiB; --auth-user requires --auth-pass.

Security

• Revisions are validated wherever they become paths (download, export, web
  API): ..-style revisions could read files outside the cache through the
  new export endpoint during this release cycle (fixed before release).

Behavior Changes

• Go 1.25 is now the minimum for building from source (security fixes in
  golang.org/x/net require it; Go 1.24 is no longer supported upstream).
  Release binaries and the Docker image are built with Go 1.26.8.
• On Windows, cache entries are hardlinks instead of missing.
• Filtered GGUF downloads also skip large LFS .imatrix files.
• list/info show sizes as KiB/MiB/GiB.

Internal

• golang.org/x/net v0.56.0 / x/sys v0.46.0: govulncheck finds no
  vulnerabilities; staticcheck clean; the codebase is gofmted.
• Docs (docs/CLI.md, docs/API.md, README) were checked flag by flag and
  endpoint by endpoint against the binary.

Full Changelog: v3.3.0...v3.4.0