HuggingFace
huggingfaceInspect HuggingFace Hub models, files, and quantization analysis quickly. Check model metadata, checkpoint sizes, and gated repo details — run structured queries by action and query type without downloading weight files.
Instructions
Inspect HuggingFace Hub models: metadata, files, and quantization quality.
Use this for model-repo lookups: architecture and parameter count, checkpoint size and shard layout, gated/private status, base-model lineage, per-file sizes and LFS checksums, and a quantization analysis that reports effective bits per weight rather than what the uploader claimed. huggingface.co URLs are also handled automatically by web_fetch_incisive — this tool is for structured queries.
Actions: model, file, tree, search, org.
Query formats vary by action:
model: "org/name" (e.g. "openai/gpt-oss-120b"), optionally "org/name@revision"
file: "org/name/path/to/file" (e.g. "openai/gpt-oss-120b/config.json") — use ref= for a branch/tag
tree: "org/name" or "org/name/subdirectory"
search: free text (Hub search is substring-based over repo ids); pair with author= to scope
org: an organization or user name (e.g. "mlx-community")
Weight files are never downloaded. Asking for a .safetensors or .gguf file returns its size, LFS checksum, and the byte-range recipe for reading the header — a multi-GB shard exposes its per-tensor dtypes and shapes in a header of a few hundred KiB.
On the model action, effective bits-per-weight is suppressed rather than guessed whenever the Hub's own numbers cannot support it: packed storage counts reported as parameter counts, repos shipping more than one checkpoint set, diffusers pipelines, and GGUF-only repos each get an explicit explanation instead of a misleading number. Set quant_audit=true to spend one extra request reading the base model's native weight format, which resolves several of those cases and yields a grid-preservation verdict.
Authentication: optional. Set HF_TOKEN env var or create ~/.config/parkour/hf_token to reach gated and private repos and raise the rate limit. Without a token the Hub returns an identical 401 for gated, private, and nonexistent repos, and this tool reports that ambiguity rather than guessing which one it hit.
Tool call template
Replace each placeholder with a value matching the input schema. Only required parameters are included; add optional parameters as needed.
{
"id": 1,
"jsonrpc": "2.0",
"method": "tools/call",
"params": {
"arguments": {
"action": "<action>",
"query": "<query>"
},
"name": "huggingface"
}
}Input Schema
| Name | Required | Description | Default |
|---|---|---|---|
| ref | No | Git revision (branch, tag, or commit SHA) for file/tree. Defaults to main. | |
| sort | No | Sort field for search/org: downloads, likes, lastModified, or trendingScore. | downloads |
| limit | No | Maximum results for search/org (default 10, max 100). | |
| query | Yes | For model/tree: 'org/name' (optionally 'org/name@revision'). For file: 'org/name/path/to/file'. For search: free text (Hub search is substring-based over repo ids). For org: the organization or user name. Any huggingface.co URL is also accepted and routed automatically. | |
| action | Yes | The operation to perform. model: model metadata, quantization analysis, and model card. file: read a repo file (weight files are described, never downloaded). tree: list repo files with sizes and LFS checksums. search: find models by name, optionally scoped to an author. org: list an organization's or user's models. | |
| author | No | Scope a search to one organization or user. | |
| quant_audit | No | On the model action, spend one extra request to read the base model's native weight format. Buys the grid-preservation verdict and a trustworthy parameter count when the Hub reported packed storage elements instead of logical weights. |