Language: English | 中文
A local model library for macOS. Download once, link everywhere.
Running LM Studio, Ollama and llama.cpp side by side means storing the same 7B model three times — 15 GB to hold 5 GB of weights. Silo keeps one physical copy in a content-addressed store and hard-links it into each tool's directory. Every tool sees an ordinary file; the disk gives up the space once.
Downloading fast from China is the reason this started, but it is not what the project is. Silo is a library manager that happens to have a good downloader.
 |
 |
 |
| One search box, every source at once — GGUF and MLX builds surfaced instead of buried. | Paste an exact author/repo and every quant is there, sized, with shards grouped into one. |
What is on disk, which tools it is linked into, and what deduplication gave back. |
- One copy on disk. A content-addressed store holds the bytes once and hard-links them into every tool that wants them — same inode, no extra space, and each tool still sees an ordinary file.
- A downloader built for slow links. Ranged parallel chunks, resume across restarts and reboots, SHA-256 verification, rate limiting.
- Mirrors that fail over. Both hubs publish the same digest for the same file, so a download can switch mirrors mid-flight and still verify.
- Picks the fastest mirror by measuring it. Not by trusting the order they are listed in — real bytes are sampled before choosing.
- One search box for every source. Keyword search across all sources at once, with the GGUF and MLX builds surfaced instead of buried.
- Sharded models stay whole. Multi-part weights and the
mmproj-*.ggufa vision model needs are grouped into one selectable variant. - A serial queue. Downloading two models at once is slower than doing them in turn, so the queue runs one at a time — and survives Ctrl-C.
- Both a window and a terminal. The macOS app and the
silocommand share the same engine and the same library.
Sources
- HuggingFace
- hf-mirror
- ModelScope
- Custom endpoints
Targets
- LM Studio
- Ollama
- llama.cpp
- ComfyUI
- HuggingFace cache (transformers / vLLM)
The app and the command come from the same tap, and it strips the quarantine attribute for you — nothing to do after installing.
brew tap linxunfeng/tap
brew install --cask linxunfeng/tap/silo # the app
brew install linxunfeng/tap/silo # the CLI--cask is what tells the two apart: the formula and the cask share a name, so
without it Homebrew installs the command.
brew upgrade --cask silo # the app
brew upgrade silo # the CLIDownload the .dmg from Releases
and drag Silo into Applications.
The builds are neither signed nor notarised, so Gatekeeper blocks the first launch. Strip the quarantine attribute to get past it:
xattr -dr com.apple.quarantine /Applications/Silo.appDownload the tarball for your architecture from the same release — uname -m
says which one is yours:
tar -xzf silo-*-macos-$(uname -m).tar.gz
xattr -d com.apple.quarantine silo 2>/dev/null || true
mv silo ~/bin/siloOr build it from source:
dart pub get
dart compile exe packages/cli/bin/silo.dart -o ~/bin/siloThe app does all of this with a window. Its one search box takes either a
keyword or an exact author/repo, and does the sensible thing with each —
search results you can click through, or straight to the variant list.
# Don't know the exact id? Search every source at once.
# Only GGUF and MLX repositories are shown; --all for the rest.
silo search qwen3 coder
# What does this repo offer, and which mirrors carry it?
silo inspect Qwen/Qwen2.5-0.5B-Instruct-GGUF
# Which mirror is actually fastest right now?
silo sources --probe Qwen/Qwen2.5-0.5B-Instruct-GGUF
# Pull it in (defaults to Q4_K_M) and hard-link it into LM Studio.
# Several models queue up and run one at a time.
silo add Qwen/Qwen2.5-0.5B-Instruct-GGUF Qwen/Qwen2.5-1.5B-Instruct-GGUF
# Ctrl-C is safe: the queue and the partial bytes both survive
silo queue --resume
# What is in the library, and what did dedup save?
silo ls
# What do my tools already have?
silo scan
# Take a model out of a tool but keep it in the store
silo unlink Qwen/Qwen2.5-0.5B-Instruct-GGUF --from lmstudio
# Forget a model (unlinks it too), then reclaim the space
silo rm Qwen/Qwen2.5-0.5B-Instruct-GGUF
silo gcUseful global flags: -j connections (default 8), --limit 5M to throttle,
--source modelscope to pin a mirror, --store to relocate the library.
Source Engine Target
────────────────── ────────────────── ──────────────────
HuggingFace range requests LM Studio
hf-mirror parallel chunks llama.cpp
ModelScope resume via sidecar Ollama
(custom endpoints) SHA-256 verification ComfyUI / HF cache
rate limiting
The engine in the middle knows nothing about models. It moves bytes from a URL into a file. Sources and targets are plugins on either side, so adding a mirror never means touching a target, and adding a tool never means touching the downloader.
packages/core engine, sources, store, targets — zero runtime dependencies
packages/cli the `silo` command
apps/silo Flutter macOS app
silo_core deliberately has no package dependencies: dart:io and a hand-rolled
streaming SHA-256, pinned against NIST vectors and the system shasum.
Version numbers come out of the commit history, so commits follow
Conventional Commits. One product, one
number: silo_core, silo_cli and the app are versioned in lockstep.
dart run melos run release # bump the workspace, write the changelogs, tag v<version>
git push --follow-tags # the tag is what builds the DMG and cuts the releaseLocal GitHub Actions
───────────────────────────────── ─────────────────────────────────
conventional commits
│
▼
melos run release
├ melos version --all one version across all three packages,
│ dependents' constraints move with it
├ CHANGELOG × 4 one per package, plus the root summary
├ chore(release) commit
└ git tag v0.1.1
│
▼
git push --follow-tags ──────────▶ on: push tags ['v*']
│
check analyze + test
│
meta version read off the tag,
│ checked against the pubspec
┌──────────┴──────────┐
cli app
arm64 + x86_64 flutter build macos
→ a tar.gz each → hdiutil → DMG
└──────────┬──────────┘
release
SHA256SUMS + gh release create
│
tap checksums →
LinXunFeng/homebrew-tap,
which opens its own bump PR
melos stops at the tag. Everything past it — the two CLI tarballs, the DMG, the checksums, the release itself, the Homebrew bump — is the workflow's job.
A CLI tarball per architecture rather than one universal binary: dart compile exe appends the AOT snapshot past the end of the Mach-O, and lipo copies only
the Mach-O, so a lipo'd silo starts up as the bare Dart VM.
The tap job needs TAP_APP_ID and TAP_APP_PRIVATE_KEY — a GitHub App with
contents and pull-request write on the tap. GITHUB_TOKEN cannot reach another
repository.
melos run analyze and melos run test are exactly what CI runs, so they are
the two commands worth running before pushing.
Apache License 2.0 © 2026 LinXunFeng
- GitHub: https://github.com/LinXunFeng
- Email: linxunfeng@yeah.net
- Blogs:
