The Hub hosts over a million models. Finding the right one means filtering down, then reading the card before you download.
1. Filter the Models page
The left sidebar on huggingface.co/models filters by task (the pipeline tag, e.g. text-generation), library (transformers, gguf, diffusers), language, license, and parameter count. Stack filters to narrow fast. Each filter is written into the URL, so a filtered search is a shareable link.
2. Read the sort signals
Sort by Trending (the default), Most downloads, or Most likes. Trending reflects recent momentum, not quality. Downloads are cumulative and include automated CI pulls, so a high count means widely used, not best. Likes are manual bookmarks, the closest thing to a human vote.
3. Read the model card
Open the card (the repo README.md) and check the license, the intended-use and limitations section, and any eval tables. The base_model field in the metadata tells you what you are looking at: a fine-tune, an adapter, a merge, or a quantized derivative of another model. A repo tagged base_model_relation: quantized is someone else’s compression of an original; verify the source.
4. Pick a file in the Files tab
The Files tab lists every file with its size. For local use you want one safetensors shard set or a single GGUF, not the whole repo. The filename encodes the quantization (Q4_K_M, fp16); match it to your hardware with RAM and VRAM requirements and the quantization guide, then grab it with the hf CLI.
5. Search from the API
from huggingface_hub import HfApi
api = HfApi()
models = api.list_models(filter=["text-generation", "gguf"], sort="downloads", limit=10)
Returns the top 10 text-generation GGUF repos by download count. The search guide documents filter, author, sort and num_parameters="min:6B,max:128B". From the shell:
hf models ls --search "llama" --sort downloads --limit 5
Lists the five most-downloaded repos matching “llama”. On any model page, the Use this model button copies a ready-made load snippet for Transformers, Ollama and other local apps.
If it fails:
- Filters return nothing — you stacked too many. Drop the size or language filter first; parameter-count filters only work on repos that declared it.
- “You need to agree to share your contact information” — the model is gated. Accept the license on the page once, or downloads return 401.
- No GGUF in Files — the repo ships only full weights. Search for a
<name>-GGUFderivative, or follow the base-model backlink.