Start
Quickstart
Find a model, download one file, and prove the bytes, in three requests and no key.
MarkdownYou want a file from a model and you want to know it is the right file. Three requests do it. Nothing to install, nothing to sign up for.
Every command on this page runs as written. hexgrad/Kokoro-82M is a real model in the index; swap in any other <owner>/<name> the search returns.
1. Find a model
curl -s "https://gethologram.ai/api/models?search=kokoro&limit=2"
[
{
"id": "hexgrad/Kokoro-82M",
"pipeline_tag": "text-to-speech",
"downloads": 12859717,
"likes": 7433,
…
},
…
]
Truncated to the fields that matter. The rows are Hugging Face's, field for field, so a model that exists there is found here.
2. List its files
curl -s https://gethologram.ai/api/models/hexgrad/Kokoro-82M/tree/main
[
{ "type": "file", "oid": "484ac8dcd971465258007528fb8bc4a99cf98633", "size": 534, "path": "EVAL.md" },
{ "type": "file", "oid": "15769c415594dcb424579ba3dfae3085578aac4e", "size": 327212226, "path": "kokoro-v1_0.pth",
"lfs": { "oid": "496dba118d1a58f5f3db2efc88dbdc216e0483fc89fe6e47ee1f2c53f18ad1e4", "size": 327212226, "pointerSize": 134 } },
…
]
This is Hugging Face's tree, unchanged. A large file's lfs.oid is the SHA-256 its bytes must have; a small file's oid is its git blob SHA-1. Keep them: they are the only hashes you will trust.
3. Fetch the file
curl -sL https://gethologram.ai/hexgrad/Kokoro-82M/resolve/main/EVAL.md -o EVAL.md
Without -L you see what the hub actually does: a 302 to a source that was up a moment ago. The weights never pass through the hub.
HTTP/1.1 302 Found
Location: https://huggingface.co/hexgrad/Kokoro-82M/resolve/f3ff3571791e39611d31c381e3a41a3af07b4987/EVAL.md
X-Hub-Source: huggingface.co
Etag: "484ac8dcd971465258007528fb8bc4a99cf98633"
4. Prove it
git hash-object --no-filters EVAL.md
484ac8dcd971465258007528fb8bc4a99cf98633
The hash on the left came from step 2, the index. The bytes came from step 3, a source. They agree, so the file is what the index says it is, and the source never had a chance to say otherwise. That is the whole product. The Verification page has the rule in full.
The same thing with one variable
Every tool built on huggingface_hub reads from the hub once you set one environment variable. Same commands, same cache.
export HF_ENDPOINT=https://gethologram.ai
hf download hexgrad/Kokoro-82M EVAL.md
transformers, diffusers, sentence-transformers, vLLM and SGLang all go through that variable; llama.cpp reads MODEL_ENDPOINT instead. The Hugging Face tools page lists each one.
Next
- Pull a whole model as a verified OCI artifact: OCI artifacts
- Run a GGUF straight into Ollama: Ollama
- Choose which source serves you: Sources and failover