GGUF model dataset

A free, regularly updated table of 215,180 GGUF models and 1,926,576 quant files: parameter counts, context lengths, file sizes and memory estimates. Download it as CSV for research, tools or your own comparisons.

models.csv 215,180 rowsfiles.csv 1,926,576 rows

What the dataset contains

Two linked tables. models.csv has one row per GGUF repository, with the model’s family, publisher, license, architecture, parameter count and trained context length. files.csv has one row per quant file, with its size and an estimate of the memory needed to run it. Join them on the model slug. The catalog was last updated on 2026-11-16.

Sample rows

ModelFamilyParametersContextQuantsFile sizes
Qwen3.8 Flash Next GSQ RCOQwen177B256K466.4 to 83.6 GB
Qwen3.8-27BQwen27.3B256K270.0 to 54.7 GB
Ornith-1.5-9BOrnith9.2B256K55.8 to 18.4 GB
Ornith-1.5-35BOrnith35.5B256K521.7 to 71.1 GB
Qwen3.8-27B iMatrix NVFP4 MTPQwen461M256K117.1 to 17.1 GB
Qwen3.8-27BQwen27.3B256K486.6 to 54.7 GB

The most downloaded language models. The full tables are in the CSV files.

models.csv columns

ColumnMeaning
slugThe model’s URL slug on ggufmodels.com
urlThe full page address
nameModel name
kindlanguage or other (image, audio and other non-language models)
familyModel family, such as Llama or Qwen
uploaderThe Hugging Face user or organization that published the GGUF files
hf_repoThe Hugging Face repository
base_modelThe original model the GGUF was made from, when listed
licenseLicense as listed on Hugging Face
architectureArchitecture from the GGUF header, such as llama or qwen2
parameters_billionTotal parameters in billions (Hugging Face’s count when available)
active_parameters_billionActive parameters of a mixture-of-experts model, when the name states it
is_mixture_of_experts1 or 0
context_lengthTrained context length in tokens, from the GGUF header
layersNumber of layers, from the GGUF header
size_classTiny, Small, Medium or Large
categoriesSemicolon-separated: General, Coding, Vision, Reasoning, Roleplay, Embedding, Image, Audio, Other
abilitiesSemicolon-separated: tools, vision, reasoning, coding, multilingual, roleplay, embedding
quant_countNumber of quant files in the repository
smallest_file_gbSize of the smallest quant file
largest_file_gbSize of the largest quant file
hf_downloadsHugging Face download count when last read
hf_likesHugging Face likes
hf_created_atWhen the repository was created
hf_updated_atWhen it was last updated

files.csv columns

ColumnMeaning
model_slugLinks the file to a row in models.csv
quantQuant type, such as Q4_K_M
filenameFile name (the first part of a split file)
size_bytesSize in bytes, all parts of a split file added up
size_gbSize in gigabytes (109 bytes)
est_memory_gb_8kEstimated memory: file size + fp16 KV cache at an 8K context + 0.5 GB runtime
partsNumber of parts the file is split into
hf_urlThe file page on Hugging Face

License and how to cite

The dataset is released under CC BY 4.0. You may copy, redistribute and adapt it, including for commercial use, as long as you give credit. Suggested citation:

ggufmodels.com. GGUF model catalog dataset. https://ggufmodels.com/data

Model files themselves are not part of the dataset and are not hosted here. Each model keeps the license listed in its repository on Hugging Face.

Limits to keep in mind

The data is read automatically from Hugging Face and from the GGUF file headers, so a repository with an unusual layout can occasionally be misread. Parameter counts use Hugging Face’s figure when it has one, otherwise our best estimate. Memory figures are estimates. Read how the data is gathered for the full method.

Frequently asked questions

Is the GGUF model dataset free to use?

Yes. It is free to download and use, including commercially, under the CC BY 4.0 license: you only have to credit ggufmodels.com. The underlying facts, such as file names, sizes and download counts, come from Hugging Face’s public API.

How often is the dataset updated?

The files are generated from the live catalog and refreshed about once an hour. The catalog itself checks for new uploads every half hour and refreshes the whole list daily.

How are the memory estimates calculated?

Each estimate is the file size plus an fp16 KV cache for an 8K context window, computed from the model’s own header, plus 0.5 GB of runtime overhead. They are estimates, not measurements.

Does the dataset include benchmark scores?

Not yet. It describes the files and their specifications: sizes, quants, context length, parameter counts and memory estimates.