HuggingFace Model Downloader 3.4.0

What's new in 3.4.0

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, …

Full release notes »