263 MB across 2 parts, newest sealed 30 October 2026 at 06:04 UTC.
The first build wrote 2 parts and 885,448 rows.
ONE_ROW_PER_MODEL passed.
The Hugging Face model repository id, publisher and name, such as Qwen/Qwen3-8B.
The account that publishes the model, the part of the id before the slash.
The publisher's avatar on the Hub.
The model's page on the Hub.
The Hub's pipeline task, such as text-generation, sentence-similarity or automatic-speech-recognition.
The library the model is published for, such as transformers, diffusers, sentence-transformers or gguf.
The architecture family from the Hub's tags, such as llama, qwen3, mistral, bert or whisper.
What the model does: instruct, base, reasoning, code, embedding, reranker, vision_language, image_generation, speech_audio, classifier, vision, other or unknown. Settled by the Hub task where it decides, by the name for text-generation models, and otherwise by a small classification model reading the card.
Where model_kind came from: hub_task, name, classifier or none.
The classifier's probability for model_kind, null when a rule settled it.
real, empty_card, test_or_toy or unlabelled. Cards under 200 characters and the unfilled template are empty_card by rule; names like tiny-random or dummy are test_or_toy by rule; a small classification model judges the rest.
The classifier's probability for card_quality, null when a rule settled it.
True when the model is a test, toy, skeleton or placeholder checkpoint.
Languages the publisher tagged on the Hub, as English names in a comma-separated list, such as English, Chinese, French.
How many languages the publisher tagged.
True when the publisher tagged the model multilingual or tagged five or more languages.
The licence tag on the Hub, such as apache-2.0, mit, llama3 or gemma, falling back to the card's metadata header.
The model this one was fine-tuned, quantized, adapted or merged from, as its Hub id.
How this model derives from base_model: finetune, quantized, adapter or merge.
How many models in this table name this model as their base, across every relation.
How many models in this table are fine-tunes of this model.
Datasets the publisher tagged the model as trained on, as Hub ids in a comma-separated list.
arXiv papers the model's tags or card cite, as ids in a comma-separated list.
Every tag on the model's Hub page, comma-separated: library, architecture, task, language codes, licence, base model, datasets and format flags.
Parameters in billions, from the safetensors or GGUF metadata, else from the size in the name.
Billions of parameters as the model name states them, such as 8 for Llama-3.1-8B.
The precision the weights are stored in: bf16, fp16, fp32, 8-bit, gguf-4bit, gguf-8bit or gguf-16bit.
Size of the weights in GB, from the parameter counts per dtype in the safetensors metadata, or for an unquantized model the name's size at 2 bytes a parameter. Null for a quantized copy with no safetensors metadata.
True when the model is a quantized copy of another: a GGUF, AWQ, GPTQ, MLX, FP8 or similar repository.
True when the repository holds GGUF files.
The smallest single GPU the weights fit with headroom: gpu_8gb, gpu_24gb (such as an RTX 4090), gpu_48gb (A6000, L40S), gpu_80gb (H100, A100) or multi_gpu. Null when the size is unknown.
True when the weights are at most 20 GB, so they fit one 24 GB consumer GPU.
True when the weights are at most 40 GB.
True when the weights are at most 70 GB, so they fit one H100 or A100 80GB.
The context length in tokens, as a number the card states. Null when the card states none.
The context length as a bucket: up_to_8k, k16_to_32k, k64_to_128k, k256_plus or not_stated. From context_tokens where known, else the classifier's reading of the card.
Where context_window came from: stated, classifier or none.
Hub downloads over the last 30 days.
Hub downloads since the repository was created.
Hub likes.
The Hub's trending score at the mirror's read.
When the repository was created, UTC.
When the repository last changed, UTC.
True when the hub-stats mirror still lists the model; false when it has since been deleted or hidden.
Whether the classification model answered for this card.
Whether the card was long enough and not the empty template, so the classifier read it.
Length of the model card in characters.
The first 4,000 characters of the model card as plain prose: metadata header, HTML, Markdown and code blocks stripped.
