unmuzzle
Agent-native distribution for open-weight models.
AI agents are the users here. unmuzzle is built so an agent (Claude Code,
Codex, Kimi Code, any MCP client) can publish and download models end to end
with no human in the loop: plain-JSON index, --json on every command, an MCP
server, and no gates, auth flows, or license click-throughs that only a human
in a browser can complete.
It is also censorship-resistant by construction:
- The index is plain JSON in a git repo. Anyone can mirror or fork it.
It also lives off-GitHub at
https://pub-95a589dcef724d6497b5ebd7ba2bade8.r2.dev/index.json(auto-synced on every change); use any copy with--indexorUNMUZZLE_INDEX. - Weights move over BitTorrent (web-seeded) and HTTP mirrors. Web seeds
keep every torrent alive at zero peers; no central server to block or bill.
Models are also on IPFS (see
mirrors.ipfsin each entry), the one channel with no company in the path: any IPFS node anywhere can fetch and re-serve the same bytes. - Every file is sha256-pinned and the manifest is minisign-signed. The signature, not the host, is the trust root. Any static host can be a mirror.
- Publisher keys are pinned on first use (TOFU). A changed key aborts the
install until you confirm the rotation with
--accept-new-key. The officialunmuzzlekey is additionally pinned inside the pip package itself, a second channel no index host can touch. - Installs land in the Hugging Face cache layout, so
transformersloads them with zero code changes.
Zero dependencies. Python 3.9+.
Install
pip install unmuzzle
# or from source: pip install git+https://github.com/zengjiajun0623/unmuzzle-hub.git
# or without PyPI at all: pip install https://pub-95a589dcef724d6497b5ebd7ba2bade8.r2.dev/unmuzzle-0.2.0-py3-none-any.whl
# optional: aria2 for torrents, minisign for signature verification
For agents
Read AGENTS.md. It is the complete publish/download protocol. Or add the MCP server:
pip install 'unmuzzle[mcp]'
claude mcp add unmuzzle -- unmuzzle-mcp
For humans
unmuzzle list
unmuzzle info unmuzzle/qwen2.5-14b-abliterated
unmuzzle get unmuzzle/qwen2.5-14b-abliterated --require-signature --dest ./models
Every command takes --json.
First release: unmuzzle/qwen2.5-14b-abliterated
Qwen2.5-14B-Instruct with the CCP-censorship refusal direction weight-baked out (GGUF Q4_K_M, ~9 GB, Apache-2.0). Answers factually on Tiananmen, Xinjiang, Taiwan, and similar topics. Runs in Ollama/llama.cpp on a 16 GB machine:
unmuzzle get unmuzzle/qwen2.5-14b-abliterated --require-signature --dest ./models
ollama create unmuzzle-qwen14b -f Modelfile # Modelfile in the HF mirror repo
ollama run unmuzzle-qwen14b
Mirrors: Hugging Face HTTP + web-seeded torrent. Signed with the unmuzzle minisign key.
Publish a model
unmuzzle keygen # one-time signing identity
unmuzzle publish ./mymodel \
--name org/mymodel \
--http-base https://your-mirror/mymodel \
--torrent-url https://your-mirror/mymodel.torrent \
--magnet "magnet:?xt=..." \
--sign-key ~/.minisign/unmuzzle.key
Then PR your index/models/*.json into this repo, or host your own index and
point clients at it with --index / $UNMUZZLE_INDEX. The full recipe,
including torrent creation with web seeds, is in AGENTS.md.
Format and threat model: SPEC.md.
Why not just Hugging Face / a mirror / a torrent site?
- HF gates models behind human browser flows and can take them down (ERNIE ViLG, GEITje). In mainland China the site itself is blocked, which is why hf-mirror.com exists. Right default for ordinary models, wrong single point of failure for models someone wants suppressed, and unusable for agents hitting gated repos.
- Mirrors (hf-mirror, ModelScope) reintroduce a single operator who can censor, and you cannot verify what they serve back to you.
- A bare torrent has no discovery, no publisher identity, and dies at zero seeds. unmuzzle adds the index, per-file sha256, publisher signatures, and web seeds on top.
Roadmap
- first real release (unmuzzle/qwen2.5-14b-abliterated)
- MCP server for agent-native publish/download
- standalone registry site generated from the index
- DHT-native peer discovery, built-in torrent client (drop aria2c)
- index federation: subscribe to multiple publisher indexes
License
MIT
Metadata
Release files for unmuzzle 0.3.0
For a detailed explanation of source distributions (sdists) and built distributions (wheels), please see the package formats documentation.
Source distribution (sdist)
| File | Size | Uploaded | Provenance | Actions |
|---|---|---|---|---|
| unmuzzle-0.3.0.tar.gz | 23.1 kB | Without provenance | Details for unmuzzle-0.3.0.tar.gz |
Built distribution (wheel)
| File | Interpreter | ABI | Platform | Actions Reset |
|---|---|---|---|---|
| unmuzzle-0.3.0-py3-none-any.whl | Python 3 | none | any | Details for unmuzzle-0.3.0-py3-none-any.whl |
Total release size: 44.3 kB
Release files / unmuzzle-0.3.0.tar.gz
| Download URL | unmuzzle-0.3.0.tar.gz |
|---|---|
| Size | 23.1 kB |
| Tags | Source |
|
SHA-256 checksum How to use checksums |
7f969671c6e08f9d6078347ad772f51e6b48217a1eb02f1ba95d518703abd174
|
|
BLAKE2b-256 checksum How to use checksums |
4abac23196f8c6d787c583c65701d0bf66e0c2f4a9acc6475b094d4b50db3a84
|
| Upload date | |
|
Uploaded using Trusted Publishing? What is trusted publishing? |
No |
| Uploaded via |
twine/6.2.0 CPython/3.14.6
|
Release files / unmuzzle-0.3.0-py3-none-any.whl
| Download URL | unmuzzle-0.3.0-py3-none-any.whl |
|---|---|
| Size | 21.2 kB |
| Tags | Python 3 |
|
SHA-256 checksum How to use checksums |
6c6d1881a9789d27ace6d07ecb932344228a19ece74c066e752a2c0780967df4
|
|
BLAKE2b-256 checksum How to use checksums |
3047483aa6e0a0f24406bc0b979545c64f929eb083164ca0b0a5995aea696322
|
| Upload date | |
|
Uploaded using Trusted Publishing? What is trusted publishing? |
No |
| Uploaded via |
twine/6.2.0 CPython/3.14.6
|
Security for unmuzzle 0.3.0
Provenance
Provenance describes where a file came from. On PyPI, provenance is shared via attestations, which provide a verifiable record of the build or publishing details. View details, limitations and caveats.
No provenance found
None of the files in this release were uploaded with provenance metadata.
A note on package security
PyPI is a repository for distributing packages, not a service for auditing them. While third-party security researchers and our staff do their best to detect and remove malicious packages, PyPI cannot guarantee the quality, security or suitability of any given package.
To learn more about protecting yourself from malicious actors, visit our package security documentation.