Datasets:

JoesSattes
/

hub-weight-census

id large_stringlengths 7 113	status large_stringclasses 7 values	downloads_30d int64 0 183M	downloads_all_time int64 0 4.2B	likes int64 0 19k	created_at large_stringdate 2022-03-02 23:29:04 2026-11-02 04:15:25	library large_stringclasses 262 values	pipeline_tag large_stringclasses 54 values	origin large_stringclasses 5 values	stratum large_stringclasses 32 values	weight float64 1 1.58k	frontier bool 2 classes	family large_stringclasses 7 values	family_config large_stringclasses 7 values	params float64 0 1,650B ⌀	hub_params float64 0 1,650B ⌀	params_match_hub bool 2 classes	storage large_stringclasses 23 values	d float64 1 16.4k ⌀	layers float64 2 162 ⌀	heads float64 1 2.3k ⌀	kv_heads float64 0.5 272 ⌀	head_dim float64 1 2.05k ⌀	ffn float64 2 1.18M ⌀	experts float64 2 512 ⌀	experts_per_token float64 1 64 ⌀	vocab float64 1 1.77M ⌀	ssm_layers float64 0 130 ⌀	kv_bytes_per_token_bf16 float64 256 17.8M ⌀	has_config bool 2 classes	config_model_type large_stringclasses 728 values	config_architectures large_stringclasses 940 values	config_d float64 0 16.4k ⌀	config_layers float64 0 162 ⌀	config_heads float64 0 128 ⌀	config_kv_heads float64 0 128 ⌀	config_vocab float64 0 902k ⌀	config_experts float64 0 2.56k ⌀	header_bytes float64 88 54.1M ⌀	weight_bytes float64 2.37k 11,827B ⌀	tensors float64 0 375k ⌀	no_weights_kind large_stringclasses 5 values	error large_stringclasses 51 values
sentence-transformers/all-MiniLM-L6-v2	ok	182,509,097	4,200,257,364	6,055	2022-03-02 23:29:05+00:00	sentence-transformers	sentence-similarity	quantized	top	1	false	encoder	encoder	22,713,728	22,713,728	true	unquantized	384	6	12	12	32	1,536	null	null	30,522	0	9,216	true	bert	BertModel	384	6	12	null	30,522	null	11,416	976,948,716	104	null	null
cross-encoder/ms-marco-MiniLM-L6-v2	ok	66,901,235	679,066,731	438	2022-03-02 23:29:05+00:00	sentence-transformers	text-ranking	quantized	top	1	false	encoder	encoder	22,714,113	22,714,113	true	unquantized	384	6	12	12	32	1,536	null	null	30,522	0	9,216	true	bert	BertForSequenceClassification	384	6	12	null	30,522	null	12,098	890,152,505	106	null	null
BAAI/bge-small-en-v1.5	ok	49,836,681	545,450,747	662	2023-09-12 05:20:55+00:00	sentence-transformers	feature-extraction	original	top	1	false	encoder	encoder	33,360,512	33,360,512	true	unquantized	384	12	12	12	32	1,536	null	null	30,522	0	18,432	true	bert	BertModel	384	12	12	null	30,522	null	22,208	401,109,346	200	null	null
openai-community/gpt2	ok	29,307,840	962,536,600	3,912	2022-03-02 23:29:04+00:00	transformers	text-generation	original	top	1	false	decoder	decoder	137,022,720	137,022,720	true	unquantized	768	12	12	12	64	3,072	null	null	50,257	0	36,864	true	gpt2	GPT2LMHeadModel	768	12	12	null	50,257	null	14,291	5,632,417,295	160	null	null
BAAI/bge-m3	ok	30,252,576	355,857,436	3,958	2024-01-27 17:07:29+00:00	sentence-transformers	sentence-similarity	original	top	1	false	encoder	encoder	567,754,752	null	null	unquantized	1,024	24	16	16	64	4,096	null	null	250,002	0	98,304	true	xlm-roberta	XLMRobertaModel	1,024	24	16	null	250,002	null	141,790	4,587,317,404	391	null	null
google-bert/bert-base-uncased	ok	29,234,461	3,281,853,805	3,060	2022-03-02 23:29:04+00:00	transformers	fill-mask	original	top	1	false	encoder	encoder	110,106,428	110,106,428	true	unquantized	768	12	12	12	64	3,072	null	null	30,522	0	36,864	true	bert	BertForMaskedLM	768	12	12	null	30,522	null	24,056	3,454,102,158	206	null	null
sentence-transformers/paraphrase-multilingual-MiniLM-L12-v2	ok	28,384,395	665,513,772	1,496	2022-03-02 23:29:05+00:00	sentence-transformers	sentence-similarity	original	top	1	false	encoder	encoder	117,654,272	117,654,272	true	unquantized	384	12	12	12	32	1,536	null	null	250,037	0	18,432	true	bert	BertModel	384	12	12	null	250,037	null	22,464	4,620,943,233	200	null	null
autogluon/chronos-2	ok	23,513,356	97,976,455	94	2025-10-06 18:52:51+00:00	chronos-forecasting	time-series-forecasting	original	top	1	false	encoder	encoder	119,477,664	119,477,664	true	unquantized	768	36	12	12	64	3,072	null	null	3,072	0	73,728	true	t5	Chronos2Model	768	12	12	null	2	null	19,816	477,940,832	170	null	null
google-t5/t5-small	ok	23,416,949	301,650,809	687	2022-03-02 23:29:04+00:00	transformers	translation	original	top	1	false	encoder-decoder	encoder-decoder	60,506,880	60,506,880	true	unquantized	512	6	8	8	64	2,048	null	null	32,128	0	24,576	true	t5	T5ForConditionalGeneration	512	6	8	null	32,128	null	15,536	2,246,973,515	132	null	null
Qwen/Qwen3-0.6B	ok	23,175,586	258,109,736	1,868	2025-04-27 03:40:08+00:00	transformers	text-generation	finetune	top	1	false	decoder	decoder	751,632,384	751,632,384	true	unquantized	1,024	28	16	8	128	3,072	null	null	151,936	0	114,688	true	qwen3	Qwen3ForCausalLM	1,024	28	16	8	151,936	null	35,560	1,519,209,243	311	null	null
sentence-transformers/all-mpnet-base-v2	ok	20,046,382	1,369,448,988	1,402	2022-03-02 23:29:05+00:00	sentence-transformers	sentence-similarity	original	top	1	false	encoder	encoder	109,486,978	109,486,978	true	unquantized	768	12	12	12	64	3,072	null	null	30,527	0	36,864	true	mpnet	MPNetForMaskedLM	768	12	12	null	30,527	null	21,904	3,824,846,935	200	null	null
google/electra-base-discriminator	legacy_pickle	17,601,325	970,264,158	218	2022-03-02 23:29:05+00:00	transformers	null	original	top	1	false	null	null	null	null	null	null	null	null	null	null	null	null	null	null	null	null	null	null	null	null	null	null	null	null	null	null	null	null	null	null	RuntimeError: not a zip-format PyTorch file (old pickle format is not supported)
openai/clip-vit-base-patch32	ok	17,564,725	710,757,874	1,272	2022-03-02 23:29:05+00:00	transformers	zero-shot-image-classification	original	top	1	false	prefix-lm	prefix-lm	151,277,440	null	null	unquantized	512	12	8	8	64	2,048	null	50	49,408	0	24,576	true	clip	CLIPModel	512	12	8	null	49,408	null	163,269	1,819,546,255	400	null	null
google/vit-base-patch16-224	ok	15,440,331	204,657,718	1,048	2022-03-02 23:29:05+00:00	transformers	image-classification	original	top	1	false	encoder	encoder	86,567,656	86,567,656	true	unquantized	768	12	12	12	64	3,072	null	null	1,000	0	36,864	true	vit	ViTForImageClassification	768	12	12	null	null	null	23,228	1,385,537,365	200	null	null
argmaxinc/whisperkit-coreml	legacy_pickle	27,332,146	107,782,889	260	2024-02-28 08:05:21+00:00	whisperkit	automatic-speech-recognition	original	top	1	false	null	null	null	null	null	null	null	null	null	null	null	null	null	null	null	null	null	null	null	null	null	null	null	null	null	null	null	null	null	null	RuntimeError: not a zip-format PyTorch file (old pickle format is not supported)
jonatasgrosman/wav2vec2-large-xlsr-53-japanese	ok	16,178,890	123,614,599	118	2022-03-02 23:29:05+00:00	transformers	automatic-speech-recognition	original	top	1	false	prefix-lm	prefix-lm	317,838,245	null	null	unquantized	1,024	24	16	16	64	4,096	null	null	2,341	0	98,304	true	wav2vec2	Wav2Vec2ForCTC	1,024	24	16	null	2,341	null	190,015	2,542,940,783	424	null	null
timm/mobilenetv3_small_100.lamb_in1k	ok	15,087,982	1,188,391,071	156	2022-12-16 05:38:36+00:00	timm	image-classification	original	top	1	false	encoder	encoder	2,555,002	2,554,968	true	unquantized	96	11	null	null	null	null	null	null	null	0	null	true	null	null	null	null	null	null	null	null	21,768	20,548,047	244	null	null
BAAI/bge-reranker-v2-m3	ok	14,444,120	162,879,247	1,285	2024-03-15 13:32:18+00:00	sentence-transformers	text-classification	original	top	1	false	encoder	encoder	567,755,777	567,755,777	true	unquantized	1,024	24	16	16	64	4,096	null	null	250,002	0	98,304	true	xlm-roberta	XLMRobertaForSequenceClassification	1,024	24	16	null	250,002	null	48,744	2,293,567,489	393	null	null
nomic-ai/nomic-embed-text-v1.5	ok	11,853,539	175,003,787	988	2024-02-10 06:32:35+00:00	sentence-transformers	sentence-similarity	original	top	1	false	encoder	encoder	136,731,648	136,731,648	true	unquantized	768	12	12	12	64	3,072	null	null	30,528	0	36,864	true	nomic_bert	NomicBertModel	768	12	12	null	30,528	null	11,576	2,215,243,354	112	null	null
hexgrad/Kokoro-82M	ok	11,366,340	138,899,285	7,325	2024-12-26 00:20:08+00:00	null	text-to-speech	finetune	top	1	false	prefix-lm	prefix-lm	81,763,410	null	null	unquantized	128	4	null	null	null	2,180	null	null	1,028	8	null	true	null	null	null	3	null	null	null	null	256,222	363,323,757	548	null	null
prithivMLmods/Gender-Classifier-Mini	ok	10,265,515	10,848,350	19	2025-03-23 14:28:14+00:00	transformers	image-classification	finetune	top	1	false	encoder	encoder	92,885,762	92,885,762	true	unquantized	768	12	12	12	64	3,072	null	null	196	0	36,864	true	siglip	SiglipForImageClassification	768	12	12	null	256,000	null	24,944	2,487,873,789	210	null	null
amazon/chronos-2	ok	9,968,412	192,793,310	563	2025-10-30 14:54:39+00:00	chronos-forecasting	time-series-forecasting	original	top	1	false	encoder	encoder	119,477,664	119,477,664	true	unquantized	768	36	12	12	64	3,072	null	null	3,072	0	73,728	true	t5	Chronos2Model	768	12	12	null	2	null	19,816	477,940,831	170	null	null
Qwen/Qwen3-14B	ok	12,988,834	43,111,159	530	2025-04-27 03:42:45+00:00	transformers	text-generation	finetune	top	1	false	decoder	decoder	14,768,307,200	14,768,307,200	true	unquantized	5,120	40	40	8	128	17,408	null	null	151,936	0	163,840	true	qwen3	Qwen3ForCausalLM	5,120	40	40	8	151,936	null	51,240	29,552,613,776	443	null	null
intfloat/multilingual-e5-small	ok	9,281,664	128,517,956	480	2023-06-30 07:31:03+00:00	sentence-transformers	sentence-similarity	original	top	1	false	encoder	encoder	117,654,272	117,654,272	true	unquantized	384	12	12	12	32	1,536	null	null	250,037	0	18,432	true	bert	BertModel	384	12	12	null	250,037	null	22,464	2,280,201,964	200	null	null
BAAI/bge-large-en-v1.5	ok	9,177,978	178,109,902	781	2023-09-12 05:20:08+00:00	sentence-transformers	feature-extraction	original	top	1	false	encoder	encoder	335,142,400	335,142,400	true	unquantized	1,024	24	16	16	64	4,096	null	null	30,522	0	98,304	true	bert	BertModel	1,024	24	16	null	30,522	null	44,968	4,019,210,262	392	null	null
Comfy-Org/Z-Image-Turbo	ok	9,521,322	53,579,153	1,010	2025-11-26 01:00:52+00:00	diffusion-single-file	null	finetune	top	1	false	diffusion	null	10,261,202,992	null	null	fp8+nvfp4	2,560	36	32	8	128	9,728	null	null	151,936	0	147,456	false	null	null	null	null	null	null	null	null	533,368	40,670,908,213	2,318	null	null
sentence-transformers/paraphrase-multilingual-mpnet-base-v2	ok	8,985,672	147,128,258	554	2022-03-02 23:29:05+00:00	sentence-transformers	sentence-similarity	original	top	1	false	encoder	encoder	278,044,162	278,044,162	true	unquantized	768	12	12	12	64	3,072	null	null	250,002	0	36,864	true	xlm-roberta	XLMRobertaModel	768	12	12	null	250,002	null	22,584	10,851,051,682	200	null	null
Qwen/Qwen3-Embedding-0.6B	ok	8,874,546	100,857,936	1,320	2025-06-03 14:25:32+00:00	sentence-transformers	feature-extraction	finetune	top	1	false	decoder	decoder	595,776,512	595,776,512	true	unquantized	1,024	28	16	8	128	3,072	null	null	151,669	0	114,688	true	qwen3	Qwen3ForCausalLM	1,024	28	16	8	151,669	null	33,392	1,207,489,041	310	null	null
Bingsu/adetailer	ok	8,771,059	361,951,495	824	2023-04-26 00:58:45+00:00	ultralytics	null	original	top	1	false	unknown	null	17,205	null	null	unquantized	null	null	null	null	null	null	null	null	null	null	null	false	null	null	null	null	null	null	null	null	241,026	328,343,242	4	null	null
FacebookAI/xlm-roberta-base	ok	8,382,998	736,802,481	967	2022-03-02 23:29:04+00:00	transformers	fill-mask	original	top	1	false	encoder	encoder	278,885,778	278,885,778	true	unquantized	768	12	12	12	64	3,072	null	null	250,002	0	36,864	true	xlm-roberta	XLMRobertaForMaskedLM	768	12	12	null	250,002	null	24,540	6,352,430,498	204	null	null
Qwen/Qwen3-4B	ok	8,417,695	93,821,375	780	2025-04-27 03:41:29+00:00	transformers	text-generation	finetune	top	1	false	decoder	decoder	4,022,468,096	4,022,468,096	true	unquantized	2,560	36	32	8	128	9,728	null	null	151,936	0	147,456	true	qwen3	Qwen3ForCausalLM	2,560	36	32	8	151,936	null	45,808	8,060,926,626	398	null	null
facebook/contriever	ok	8,280,816	197,590,453	132	2022-03-02 23:29:05+00:00	transformers	null	original	top	1	false	encoder	encoder	109,482,752	null	null	unquantized	768	12	12	12	64	3,072	null	null	30,522	0	36,864	true	bert	Contriever	768	12	12	null	30,522	null	114,745	438,708,755	200	null	null
google/gemma-4-31B-it	ok	8,147,734	72,426,611	4,270	2026-03-11 18:22:36+00:00	transformers	image-text-to-text	finetune	top	1	false	prefix-lm	prefix-lm	31,273,088,876	31,273,088,876	true	unquantized	5,376	60	32	16	256	21,504	null	null	262,144	0	983,040	true	gemma4	Gemma4ForConditionalGeneration	5,376	60	32	16	262,144	null	160,496	62,578,686,256	1,188	null	null
BAAI/bge-base-en-v1.5	ok	8,054,219	584,427,390	540	2023-09-11 15:04:22+00:00	sentence-transformers	feature-extraction	original	top	1	false	encoder	encoder	109,482,752	109,482,752	true	unquantized	768	12	12	12	64	3,072	null	null	30,522	0	36,864	true	bert	BertModel	768	12	12	null	30,522	null	22,456	1,312,805,365	200	null	null
Comfy-Org/Qwen-Image-2.1	ok	10,607,935	16,860,819	1,246	2026-09-15 19:41:09+00:00	diffusion-single-file	text-to-image	finetune	top	1	false	prefix-lm	null	22,287,450,740	null	null	int8	4,096	36	48	48	128	null	null	null	151,936	0	1,769,472	false	null	null	null	null	null	null	null	null	1,829,344	122,840,171,411	4,798	null	null
Qwen/Qwen3-8B	ok	7,540,663	136,299,739	2,163	2025-04-27 03:42:21+00:00	transformers	text-generation	finetune	top	1	false	decoder	decoder	8,190,735,360	8,190,735,360	true	unquantized	4,096	36	32	8	128	12,288	null	null	151,936	0	147,456	true	qwen3	Qwen3ForCausalLM	4,096	36	32	8	151,936	null	46,056	16,397,461,266	399	null	null
Serveurperso/YuE2-GGUF	no_weights	7,537,625	8,847,449	22	2026-09-12 19:35:19+00:00	null	null	quantized	top	1	false	null	null	null	null	null	null	null	null	null	null	null	null	null	null	null	null	null	null	null	null	null	null	null	null	null	null	null	null	null	gguf_only	null
ISTA-DASLab/Qwen3.8-Flash-Next-GSQ-RCO-GGUF	no_weights	7,426,020	9,259,619	1,140	2026-09-07 09:41:37+00:00	gguf	image-text-to-text	quantized	top	1	false	null	null	null	null	null	null	null	null	null	null	null	null	null	null	null	null	null	null	null	null	null	null	null	null	null	null	null	null	null	gguf_only	null
FacebookAI/roberta-large	ok	6,932,811	667,143,335	346	2022-03-02 23:29:04+00:00	transformers	fill-mask	original	top	1	false	encoder	encoder	355,412,571	355,412,571	true	unquantized	1,024	24	16	16	64	4,096	null	null	50,265	0	98,304	true	roberta	RobertaForMaskedLM	1,024	24	16	null	50,265	null	48,139	7,534,491,787	395	null	null
pyannote/speaker-diarization-3.1	no_weights	6,918,404	357,863,051	4,828	2023-11-16 08:19:01+00:00	pyannote-audio	automatic-speech-recognition	original	top	1	false	null	null	null	null	null	null	null	null	null	null	null	null	null	null	null	null	null	null	null	null	null	null	null	null	null	null	null	null	null	no_weight_files	null
Qwen/Qwen3.8-27B	ok	7,915,637	20,706,092	18,969	2026-08-05 08:22:59+00:00	transformers	image-text-to-text	original	top	1	true	prefix-lm	prefix-lm	27,781,427,952	27,781,427,952	true	unquantized	5,120	64	24	4	256	17,408	null	null	248,320	48	65,536	true	qwen3_5	Qwen3_5ForConditionalGeneration	5,120	64	24	4	248,320	null	150,872	55,586,114,863	1,199	null	null
Comfy-Org/MiniMax-H3	ok	31,098,855	79,068,221	2,481	2026-07-30 22:36:12+00:00	diffusion-single-file	null	finetune	top	1	false	prefix-lm	null	51,196,741,883	null	null	int8+nvfp4	5,120	50	64	8	128	25,600	null	null	151,936	0	204,800	false	null	null	null	null	null	null	null	null	2,385,888	534,682,333,162	5,723	null	null
ibm-granite/granite-embedding-small-english-r2	ok	6,776,901	33,413,479	113	2025-07-17 20:41:53+00:00	sentence-transformers	feature-extraction	original	top	1	false	encoder	encoder	47,662,464	47,662,464	true	unquantized	384	12	12	12	32	1,536	null	null	50,368	0	18,432	true	modernbert	ModernBertModel	384	12	12	null	50,368	null	7,120	194,286,691	74	null	null
Qwen/Qwen3-VL-8B-Instruct	ok	6,915,010	78,692,156	1,236	2025-10-11 07:23:39+00:00	transformers	image-text-to-text	original	top	1	false	prefix-lm	prefix-lm	8,767,123,696	8,767,123,696	true	unquantized	4,096	36	32	8	128	12,288	null	null	151,936	0	147,456	true	qwen3_vl	Qwen3VLForConditionalGeneration	4,096	36	32	8	151,936	null	92,120	17,545,915,883	750	null	null
intfloat/multilingual-e5-large	ok	6,767,471	119,424,487	1,282	2023-06-30 07:38:19+00:00	sentence-transformers	feature-extraction	original	top	1	false	encoder	encoder	559,890,946	559,890,946	true	unquantized	1,024	24	16	16	64	4,096	null	null	250,002	0	98,304	true	xlm-roberta	XLMRobertaModel	1,024	24	16	null	250,002	null	45,528	9,557,699,267	392	null	null
laion/clap-htsat-fused	ok	6,415,448	216,641,944	176	2023-02-16 20:45:11+00:00	transformers	audio-classification	original	top	1	false	encoder	encoder	153,558,323	153,558,316	true	unquantized	768	12	12	12	64	3,072	null	50	50,265	0	36,864	true	clap	ClapModel	768	12	12	null	50,265	null	62,120	1,232,468,802	477	null	null
Qwen/Qwen3.5-9B	ok	6,412,961	74,105,741	2,291	2026-02-27 12:58:26+00:00	transformers	image-text-to-text	finetune	top	1	false	prefix-lm	prefix-lm	9,653,104,368	9,653,104,368	true	unquantized	4,096	32	16	4	256	12,288	null	null	248,320	24	32,768	true	qwen3_5	Qwen3_5ForConditionalGeneration	4,096	32	16	4	248,320	null	94,464	19,329,393,661	775	null	null
Qwen/Qwen3.5-4B	ok	6,376,702	56,404,655	1,180	2026-02-27 14:45:03+00:00	transformers	image-text-to-text	finetune	top	1	false	prefix-lm	prefix-lm	4,659,865,088	4,659,865,088	true	unquantized	2,560	32	16	4	256	9,216	null	null	248,320	24	32,768	true	qwen3_5	Qwen3_5ForConditionalGeneration	2,560	32	16	4	248,320	null	90,240	9,342,907,469	738	null	null
FacebookAI/roberta-base	ok	6,299,171	662,792,559	700	2022-03-02 23:29:04+00:00	transformers	fill-mask	original	top	1	false	encoder	encoder	124,697,947	124,697,947	true	unquantized	768	12	12	12	64	3,072	null	null	50,265	0	36,864	true	roberta	RobertaForMaskedLM	768	12	12	null	50,265	null	24,210	2,815,192,007	203	null	null
Comfy-Org/Wan-2.2	ok	9,326,230	98,314,213	989	2025-07-27 06:56:12+00:00	diffusion-single-file	image-to-video	finetune	top	1	false	encoder-decoder	null	23,725,717,082	null	null	unquantized	4,096	24	32	32	128	10,240	null	null	256,384	0	393,216	false	null	null	null	null	null	null	null	null	4,988,936	728,658,155,653	2,859	null	null
amazon/chronos-bolt-small	ok	6,127,478	46,395,160	67	2024-11-25 09:46:10+00:00	chronos-forecasting	time-series-forecasting	original	top	1	false	encoder-decoder	encoder-decoder	47,718,016	47,718,016	true	unquantized	512	6	8	8	64	2,048	null	null	2,048	0	24,576	true	t5	ChronosBoltModelForForecasting	512	6	8	null	2	null	16,760	190,900,126	143	null	null
distilbert/distilbert-base-uncased	ok	6,095,028	697,819,956	1,141	2022-03-02 23:29:04+00:00	transformers	fill-mask	original	top	1	false	encoder	encoder	66,985,530	66,985,530	true	unquantized	768	6	12	12	64	3,072	null	null	30,522	0	18,432	true	distilbert	DistilBertForMaskedLM	768	6	12	null	30,522	null	12,648	1,529,742,866	105	null	null
unsloth/Qwen3.8-27B-GGUF	no_weights	6,083,526	22,671,652	5,467	2026-08-13 08:28:40+00:00	null	null	quantized	top	1	false	null	null	null	null	null	null	null	null	null	null	null	null	null	null	null	null	null	null	null	null	null	null	null	null	null	null	null	null	null	gguf_only	null
autogluon/chronos-bolt-small	ok	6,019,165	162,682,784	86	2024-11-13 13:28:57+00:00	null	time-series-forecasting	original	top	1	false	encoder-decoder	encoder-decoder	47,718,016	47,718,016	true	unquantized	512	6	8	8	64	2,048	null	null	2,048	0	24,576	true	t5	ChronosBoltModelForForecasting	512	6	8	null	2	null	16,760	190,898,352	143	null	null
openai/clip-vit-large-patch14	ok	6,006,155	1,201,492,076	2,136	2022-03-02 23:29:05+00:00	transformers	zero-shot-image-classification	original	top	1	false	prefix-lm	prefix-lm	427,616,847	427,616,846	true	unquantized	768	12	12	12	64	3,072	null	50	49,408	0	36,864	true	clip	CLIPModel	768	12	12	null	49,408	null	71,856	6,846,537,784	592	null	null
microsoft/Florence-2-base	ok	5,985,608	30,942,187	433	2024-06-15 00:57:24+00:00	transformers	image-text-to-text	original	top	1	false	encoder-decoder	encoder-decoder	231,567,705	231,567,705	true	unquantized	768	6	12	12	64	3,072	null	null	51,289	0	18,432	true	florence2	Florence2ForConditionalGeneration	768	6	12	null	51,289	null	85,856	930,315,210	666	null	null
coqui/XTTS-v2	ok	5,965,681	132,667,274	3,929	2023-10-31 10:11:33+00:00	coqui	text-to-speech	original	top	1	false	prefix-lm	prefix-lm	466,900,598	null	null	unquantized	1,024	30	16	16	64	4,096	null	null	6,681	0	122,880	true	null	null	null	null	null	null	null	null	327,336	2,089,347,238	963	null	null
google/gemma-4-26B-A4B-it	ok	5,944,935	74,528,422	1,726	2026-03-11 21:25:57+00:00	transformers	image-text-to-text	finetune	top	1	false	prefix-lm	prefix-lm	25,805,936,206	25,805,936,206	true	unquantized	2,816	30	16	8	256	2,112	128	null	262,144	0	245,760	true	gemma4	Gemma4ForConditionalGeneration	2,816	30	16	8	262,144	128	137,504	51,644,340,088	1,013	null	null
audio-cpp/audio.cpp-gguf	ok	8,003,097	15,986,165	295	2026-07-14 05:56:32+00:00	audio.cpp	text-to-speech	quantized	top	1	false	unknown	null	3,391,506	null	null	unquantized	null	null	null	null	null	null	null	null	null	null	null	false	null	null	null	null	null	null	null	null	62,448	808,266,095,401	30	null	null
openai/gpt-oss-20b	ok	5,939,725	107,425,867	5,262	2025-08-04 22:33:29+00:00	transformers	text-generation	original	top	1	false	decoder	decoder	20,914,757,184	20,914,757,184	true	mxfp4	2,880	24	64	8	64	null	32	4	201,088	0	49,152	true	gpt_oss	GptOssForCausalLM	2,880	24	64	8	201,088	32	52,136	41,301,465,516	459	null	null
Qwen/Qwen2.5-7B-Instruct	ok	5,896,682	192,569,226	2,483	2024-09-16 11:55:40+00:00	transformers	text-generation	finetune	top	1	false	decoder	decoder	7,615,616,512	7,615,616,512	true	unquantized	3,584	28	28	4	128	18,944	null	null	152,064	0	57,344	true	qwen2	Qwen2ForCausalLM	3,584	28	28	4	152,064	null	38,864	15,242,807,270	339	null	null
pyannote/wespeaker-voxceleb-resnet34-LM	ok	5,815,981	405,446,734	218	2023-11-13 15:32:31+00:00	pyannote-audio	null	original	top	1	false	encoder	null	6,642,884	null	null	unquantized	128	13	null	null	null	null	null	null	null	0	null	false	null	null	null	null	null	null	null	null	111,151	26,650,477	218	null	null
Qwen/Qwen2.5-VL-7B-Instruct	ok	5,886,476	108,837,296	1,762	2025-01-26 09:26:37+00:00	transformers	image-text-to-text	original	top	1	false	prefix-lm	prefix-lm	8,292,166,656	8,292,166,656	true	unquantized	3,584	28	28	4	128	18,944	null	null	152,064	0	57,344	true	qwen2_5_vl	Qwen2_5_VLForConditionalGeneration	3,584	28	28	4	152,064	null	81,248	16,595,981,281	729	null	null
openai/whisper-large-v3-turbo	ok	5,664,908	126,560,599	3,496	2024-10-01 07:39:28+00:00	transformers	automatic-speech-recognition	finetune	top	1	false	encoder-decoder	encoder-decoder	808,878,080	808,878,080	true	unquantized	1,280	4	20	20	64	5,120	null	null	51,866	0	20,480	true	whisper	WhisperForConditionalGeneration	1,280	32	20	null	51,866	null	68,704	1,622,466,054	587	null	null
ornith-ai/Ornith-1.5-9B-GGUF	no_weights	5,435,094	13,789,963	621	2026-08-19 00:02:44+00:00	transformers	text-generation	original	top	1	false	null	null	null	null	null	null	null	null	null	null	null	null	null	null	null	null	null	null	null	null	null	null	null	null	null	null	null	null	null	gguf_only	null
Systran/faster-whisper-small	legacy_pickle	5,356,851	24,474,037	84	2023-11-23 09:53:51+00:00	ctranslate2	automatic-speech-recognition	original	top	1	false	null	null	null	null	null	null	null	null	null	null	null	null	null	null	null	null	null	null	null	null	null	null	null	null	null	null	null	null	null	null	RuntimeError: not a zip-format PyTorch file (old pickle format is not supported)
SamLowe/roberta-base-go_emotions	ok	5,258,012	106,375,029	717	2022-09-15 13:04:21+00:00	transformers	text-classification	original	top	1	false	encoder	encoder	124,667,678	124,667,678	true	unquantized	768	12	12	12	64	3,072	null	null	50,265	0	36,864	true	roberta	RobertaForSequenceClassification	768	12	12	null	50,265	null	24,236	1,000,820,716	202	null	null
microsoft/mdeberta-v3-base	ok	5,357,805	60,875,164	261	2022-03-02 23:29:05+00:00	transformers	fill-mask	original	top	1	false	decoder	decoder	473,209,209	null	null	unquantized	768	12	12	12	64	3,072	null	null	251,000	0	36,864	true	deberta-v2	null	768	12	12	null	251,000	null	106,965	3,399,556,957	212	null	null
intfloat/multilingual-e5-base	ok	5,160,583	73,896,828	419	2023-05-19 10:26:40+00:00	sentence-transformers	sentence-similarity	original	top	1	false	encoder	encoder	278,044,162	278,044,162	true	unquantized	768	12	12	12	64	3,072	null	null	250,002	0	36,864	true	xlm-roberta	XLMRobertaModel	768	12	12	null	250,002	null	22,584	5,322,810,412	200	null	null
pyannote/segmentation-3.0	ok	5,119,817	406,510,051	2,482	2023-09-22 12:03:10+00:00	pyannote-audio	voice-activity-detection	original	top	1	false	unknown	null	1,473,515	null	null	unquantized	null	null	null	null	null	null	null	null	null	null	null	false	null	null	null	null	null	null	null	null	74,408	5,982,351	54	null	null
pyannote/speaker-diarization-community-1	gated	5,102,899	37,551,741	3,691	2025-04-15 21:31:35+00:00	pyannote-audio	automatic-speech-recognition	original	top	1	false	null	null	null	null	null	null	null	null	null	null	null	null	null	null	null	null	null	null	null	null	null	null	null	null	null	null	null	null	null	null	RuntimeError: gated repository: accept its license on Hugging Face first
Qwen/Qwen2.5-0.5B-Instruct	ok	4,872,509	77,125,364	724	2024-09-16 11:52:46+00:00	transformers	text-generation	finetune	top	1	false	decoder	decoder	494,032,768	494,032,768	true	unquantized	896	24	14	2	64	4,864	null	null	151,936	0	12,288	true	qwen2	Qwen2ForCausalLM	896	24	14	2	151,936	null	32,288	999,604,126	290	null	null
MiniMaxAI/MiniMax-H3	ok	6,140,874	15,690,081	6,509	2026-07-28 10:45:18+00:00	minimax-h3	image-text-to-video	original	top	1	true	transducer	null	69,235,578,529	33,122,992,896	false	unquantized	5,120	64	64	8	128	25,600	null	null	151,936	0	262,144	false	null	null	null	null	null	null	null	null	486,968	498,474,749,480	3,486	null	null
zai-org/GLM-5.3-Flash	ok	6,313,299	12,573,351	3,019	2026-08-25 06:43:14+00:00	transformers	image-text-to-text	original	top	1	true	prefix-lm	prefix-lm	321,323,031,390	321,323,031,390	true	fp8	4,096	46	64	64	256	8,192	288	8	154,880	0	23,296	true	glm5_next	Glm5NextForConditionalGeneration	4,096	45	64	64	154,880	288	10,684,096	328,366,173,639	76,108	null	null
Qwen/Qwen2.5-1.5B-Instruct	ok	4,569,655	226,605,358	924	2024-09-17 14:10:29+00:00	transformers	text-generation	finetune	top	1	false	decoder	decoder	1,543,714,304	1,543,714,304	true	unquantized	1,536	28	12	2	128	8,960	null	null	151,936	0	28,672	true	qwen2	Qwen2ForCausalLM	1,536	28	12	2	151,936	null	38,536	3,098,973,447	338	null	null
dphn/dolphin-2.9.1-yi-1.5-34b	ok	4,771,746	84,774,201	90	2024-05-18 04:50:56+00:00	transformers	text-generation	finetune	top	1	false	decoder	decoder	34,388,917,248	34,388,917,248	true	unquantized	7,168	60	56	8	128	20,480	null	null	64,000	0	245,760	true	llama	LlamaForCausalLM	7,168	60	56	8	64,000	null	63,536	68,782,549,362	543	null	null
Xenova/all-MiniLM-L6-v2	no_weights	4,517,843	34,635,058	178	2023-05-02 22:46:15+00:00	transformers.js	feature-extraction	quantized	top	1	false	null	null	null	null	null	null	null	null	null	null	null	null	null	null	null	null	null	null	null	null	null	null	null	null	null	null	null	null	null	onnx_only	null
stabilityai/stable-diffusion-xl-base-1.0	ok	4,493,100	130,329,554	8,426	2023-07-25 13:25:51+00:00	diffusers	text-to-image	original	top	1	false	encoder-decoder	null	3,468,837,867	2,567,463,684	false	unquantized	1,280	32	20	20	64	5,120	null	null	49,408	0	163,840	false	null	null	null	null	null	null	null	null	336,968	76,912,765,291	2,641	null	null
facebook/opt-125m	ok	4,483,457	221,827,420	333	2022-05-11 08:25:17+00:00	transformers	text-generation	original	top	1	false	decoder	decoder	163,848,192	null	null	unquantized	768	12	12	12	64	3,072	null	null	50,272	0	36,864	true	opt	OPTForCausalLM	768	12	12	null	50,272	null	102,482	753,111,181	197	null	null
Qwen/Qwen3.6-35B-A3B	ok	4,555,162	34,531,001	3,067	2026-04-15 05:59:19+00:00	transformers	image-text-to-text	original	top	1	true	prefix-lm	prefix-lm	35,951,822,704	35,951,822,704	true	unquantized	2,048	40	16	2	256	null	256	8	248,320	30	20,480	true	qwen3_5_moe	Qwen3_5MoeForConditionalGeneration	2,048	40	16	2	248,320	256	131,368	71,926,865,825	1,045	null	null
Comfy-Org/MOSS-Image	ok	4,467,577	4,467,577	574	2026-10-16 12:42:32+00:00	diffusion-single-file	null	finetune	top	1	false	prefix-lm	null	12,015,238,876	null	null	int8	4,096	32	48	48	128	null	null	null	248,320	24	983,040	false	null	null	null	null	null	null	null	null	806,424	68,750,918,379	2,975	null	null
meta-llama/Llama-3.1-8B-Instruct	ok	4,463,539	192,376,330	8,748	2024-07-18 08:56:00+00:00	transformers	text-generation	finetune	top	1	false	decoder	decoder	8,030,261,248	8,030,261,248	true	unquantized	4,096	32	32	8	128	14,336	null	null	128,256	0	131,072	true	llama	LlamaForCausalLM	4,096	32	32	8	128,256	null	33,880	32,132,582,323	291	null	null
laion/CLIP-ViT-B-32-laion2B-s34B-b79K	ok	4,401,388	123,006,355	157	2022-09-14 22:49:28+00:00	open_clip	zero-shot-image-classification	original	top	1	false	prefix-lm	prefix-lm	302,554,753	151,277,439	false	unquantized	512	12	8	8	64	2,048	null	50	49,408	0	24,576	true	clip	CLIPModel	512	12	8	null	49,408	null	81,680	2,424,386,679	702	null	null
Qwen/Qwen3-1.7B	ok	4,205,972	67,602,856	610	2025-04-27 03:41:05+00:00	transformers	text-generation	finetune	top	1	false	decoder	decoder	2,031,739,904	2,031,739,904	true	unquantized	2,048	28	16	8	128	6,144	null	null	151,936	0	114,688	true	qwen3	Qwen3ForCausalLM	2,048	28	16	8	151,936	null	35,784	4,079,450,110	311	null	null
baidu/Unlimited-OCR	ok	4,241,405	12,765,657	4,939	2026-06-19 09:40:33+00:00	transformers	image-text-to-text	original	top	1	true	prefix-lm	prefix-lm	3,336,106,240	3,336,106,240	true	unquantized	1,280	12	10	10	128	6,848	64	6	129,280	0	61,440	true	unlimited_ocr	UnlimitedOCRForCausalLM	1,280	12	10	10	129,280	64	669,280	13,450,917,926	2,710	null	null
openai/whisper-small	ok	4,166,276	140,883,700	645	2022-09-26 06:51:27+00:00	transformers	automatic-speech-recognition	original	top	1	false	encoder-decoder	encoder-decoder	241,734,912	241,734,912	true	unquantized	768	12	12	12	64	3,072	null	null	51,865	0	36,864	true	whisper	WhisperForConditionalGeneration	768	12	12	null	51,865	null	55,432	3,872,928,558	479	null	null
Qwen/Qwen3.8-27B-FP8	ok	4,731,404	16,036,405	955	2026-08-13 08:01:58+00:00	transformers	image-text-to-text	quantized	top	1	false	prefix-lm	prefix-lm	27,781,427,952	27,781,427,952	true	fp8	5,120	64	24	4	256	17,408	null	null	248,320	48	65,536	true	qwen3_5	Qwen3_5ForConditionalGeneration	5,120	64	24	4	248,320	null	203,664	30,890,049,597	1,606	null	null
google/gemma-4-E4B-it	ok	4,128,512	38,159,835	1,819	2026-03-02 19:57:40+00:00	transformers	any-to-any	finetune	top	1	false	prefix-lm	prefix-lm	7,996,157,418	7,996,156,490	true	unquantized	2,560	42	8	2	256	10,240	null	null	262,144	0	86,016	true	gemma4	Gemma4ForConditionalGeneration	2,560	42	8	2	262,144	null	281,048	16,024,823,729	2,130	null	null
ornith-ai/Ornith-1.5-35B-A3B-GGUF	no_weights	4,036,317	10,648,283	571	2026-08-18 06:38:41+00:00	transformers	text-generation	original	top	1	false	null	null	null	null	null	null	null	null	null	null	null	null	null	null	null	null	null	null	null	null	null	null	null	null	null	null	null	null	null	gguf_only	null
Qwen/Qwen3-VL-4B-Instruct	ok	3,936,514	32,849,793	523	2025-10-11 07:23:22+00:00	transformers	image-text-to-text	original	top	1	false	prefix-lm	prefix-lm	4,437,815,808	4,437,815,808	true	unquantized	2,560	36	32	8	128	9,728	null	null	151,936	0	147,456	true	qwen3_vl	Qwen3VLForConditionalGeneration	2,560	36	32	8	151,936	null	87,728	8,887,292,732	713	null	null
Comfy-Org/Wan-2.1	ok	4,667,254	64,232,389	1,039	2025-02-25 21:27:12+00:00	diffusion-single-file	null	finetune	top	1	false	encoder-decoder	null	31,033,128,525	null	null	unquantized	4,096	24	32	32	128	10,240	null	null	256,384	0	393,216	false	null	null	null	null	null	null	null	null	3,567,684	487,425,919,360	4,098	null	null
facebook/esm2_t33_650M_UR50D	ok	3,928,554	57,239,155	108	2022-09-27 14:36:16+00:00	transformers	fill-mask	original	top	1	false	encoder	encoder	652,358,616	652,358,616	true	unquantized	1,280	33	20	20	64	5,120	null	null	null	0	168,960	true	esm	EsmForMaskedLM	1,280	33	20	null	33	null	67,824	7,824,242,163	573	null	null
datasocietyco/bge-base-en-v1.5-course-recommender-v5	ok	3,861,303	19,104,539	24	2025-01-09 11:01:33+00:00	sentence-transformers	sentence-similarity	finetune	top	1	false	encoder	encoder	109,482,240	109,482,240	true	unquantized	768	12	12	12	64	3,072	null	null	30,522	0	36,864	true	bert	BertModel	768	12	12	null	30,522	null	22,368	438,932,581	199	null	null
jonatasgrosman/wav2vec2-large-xlsr-53-russian	ok	3,840,804	144,514,060	89	2022-03-02 23:29:05+00:00	transformers	automatic-speech-recognition	original	top	1	false	encoder	encoder	315,478,695	null	null	unquantized	1,024	24	16	16	64	4,096	null	null	39	0	98,304	true	wav2vec2	Wav2Vec2ForCTC	1,024	24	16	null	39	null	190,010	4,047,623,728	424	null	null
Comfy-Org/Krea-2	ok	4,143,236	18,000,690	780	2026-06-16 22:31:20+00:00	diffusion-single-file	text-to-image	finetune	top	1	false	prefix-lm	null	17,806,108,639	null	null	fp8+nvfp4	2,560	36	32	8	128	9,728	null	null	151,936	0	147,456	false	null	null	null	null	null	null	null	null	1,770,680	164,322,444,400	2,891	null	null
ProsusAI/finbert	ok	3,705,594	126,287,003	1,296	2022-03-02 23:29:04+00:00	transformers	text-classification	original	top	1	false	encoder	encoder	109,485,059	109,485,059	true	unquantized	768	12	12	12	64	3,072	null	null	30,522	0	36,864	true	bert	BertForSequenceClassification	768	12	12	null	30,522	null	23,624	1,752,337,930	202	null	null
Qwen/Qwen3.5-0.8B	ok	3,693,505	23,257,780	793	2026-02-28 23:57:01+00:00	transformers	image-text-to-text	finetune	top	1	false	prefix-lm	prefix-lm	873,438,784	873,438,784	true	unquantized	1,024	24	8	2	256	3,584	null	null	248,320	18	12,288	true	qwen3_5	Qwen3_5ForConditionalGeneration	1,024	24	8	2	248,320	null	59,848	1,769,980,465	488	null	null
jonatasgrosman/wav2vec2-large-xlsr-53-portuguese	ok	3,665,873	184,414,223	72	2022-03-02 23:29:05+00:00	transformers	automatic-speech-recognition	original	top	1	false	encoder	encoder	315,485,870	null	null	unquantized	1,024	24	16	16	64	4,096	null	null	46	0	98,304	true	wav2vec2	Wav2Vec2ForCTC	1,024	24	16	null	46	null	190,012	3,715,719,168	424	null	null
nvidia/Qwen3.6-35B-A3B-NVFP4	ok	3,788,274	40,934,747	673	2026-05-27 18:09:46+00:00	Model Optimizer	text-generation	quantized	top	1	false	prefix-lm	prefix-lm	35,951,822,704	18,683,860,336	false	fp8+nvfp4	2,048	40	16	2	256	null	256	8	248,320	30	20,480	true	qwen3_5_moe	Qwen3_5MoeForConditionalGeneration	2,048	40	16	2	248,320	256	16,757,464	23,462,477,857	124,468	null	null
cross-encoder/ms-marco-MiniLM-L4-v2	ok	3,554,736	73,005,632	52	2022-03-02 23:29:05+00:00	sentence-transformers	text-ranking	quantized	top	1	false	encoder	encoder	19,165,185	19,165,185	true	unquantized	384	4	12	12	32	1,536	null	null	30,522	0	6,144	true	bert	BertForSequenceClassification	384	4	12	null	30,522	null	8,398	750,743,773	74	null	null

End of preview. Expand in Data Studio

Hub Weight Census

What the weights on the Hugging Face Hub say about themselves. For each sampled model repository we read only the safetensors headers (a median of 75 KB per repo, about 0.0026% of the weight bytes) and recover the architecture from tensor names, shapes and dtypes with ai-model-explorer, then compare it with the repo's config.json and with the Hub's own parameter count.

Method: S. Singkul, Reading the Architecture off the Weights: Manifest-Only Recovery of Neural Network Anatomy, 2026 (paper). Explore it on the web: https://joessattes.github.io/ai-model-explorer/census/

Citation. If you use this dataset, please cite the paper (PDF, ResearchGate); BibTeX under Citation.

Sample (2026-11-03)

frame: 3206528
sample: 12974
seed: 2026
top: 10000
rest_target: 2500
strata: 32
frontier_since: 2025-11-01
frontier_repos: 743
weight_sum: 3206998

Strata: top (most downloaded, read in full), original releases (no base_model: derivative tag, read in full), frontier (originals from leading providers created since 2025-11-01), and a random sample of the rest of the Hub stratified by creation year x download tier x safetensors tag, with weight = stratum size / sampled. Weighted estimates describe the whole Hub; the top/original/frontier strata describe the most downloaded models.

Files

census.parquet: one flat row per repo (shown in the viewer above): status, popularity, family, parameters, the Hub's count, storage format, width, layers, heads, KV heads, experts, vocabulary, KV bytes per token, and the values config.json declares.
census.jsonl: the full records, fields below.
census_manifests.jsonl.gz: the raw manifest and config.json of every readable repo (see below).
roles.json: per-tensor role accuracy on repos whose model type is in the companion corpus.
sample.parquet: the sampled repos with stratum and weight; sample_info.json: sample sizes and settings.
frame_stats.json: statistics of the full Hub frame the sample was drawn from.
report/: summary tables and charts; insights/: ecosystem findings (ecosystem.json, insights.md) and figures.

Fields (census.jsonl)

id, status (ok, gated, no_weights + no_weights_kind, legacy_pickle, not_found, timeout, error), popularity (downloads 30-day, downloads_all_time, likes), library, pipeline_tag, origin, stratum, weight; for readable repos: family_manifest (weights only), family_config, params, stored_bytes, hub_params, quant, dtypes, hp (weights only), hp_fused, config_hp (as declared), components, stacks, config_checked / config_consistent with the widths behind any mismatch, header_bytes, weight_bytes, timings.

census_manifests.jsonl.gz: for every readable repo, the raw manifest (name, dtype, shape of every tensor) and its config.json, so any field can be recomputed (rescore.py). roles.json: per-tensor role accuracy on repos whose model type is in the companion corpus (manifest-anatomy-corpus), with labels carried over by matching tensor names after layer indices are masked.

Datasheet

Motivation. Measure how much of the public model ecosystem can be described from weight metadata alone, how often no other specification exists, and where the Hub's own metadata is wrong.
Collection. Public Hub API and HTTP range requests on safetensors headers; no weights downloaded, no gated content read (gated repos are recorded as such, with their public popularity stats only).
Frame. cfahlgren1/hub-stats models snapshot from the run date.
Known limits. config_hp is the reference, and configs can themselves be wrong; head counts need a config (shapes give heads x head size only); MLX bit-width and group size trade off without a config; repos stored only in the pre-1.6 PyTorch pickle format are not parsed.
Personal data. None. Repository ids and authors are public account names.
License. MIT for this dataset; the models described keep their own licenses.

Findings

From 9,939 popular repos (most downloaded, original releases, recent frontier releases) out of 12,974 read.

Mixture of experts went from 0% of popular original language models released in 2022 to 42% in 2026.
Hybrid stacks (state-space or linear-attention layers next to attention): 0% → 28%.
Grouped-query attention (fewer KV heads than query heads): 0% → 84%.
Median head size 64 → 128; median vocabulary 50,257 → 151,936 tokens.
Original releases published already quantized (FP8, NVFP4, MXFP4 …): 0% → 24%.
The Hub's parameter count is wrong for packed formats: nvfp4 89% of 70 repos (true/Hub median 1.758x); fp8+nvfp4 100% of 39 repos (true/Hub median 1.539x).
Per-tensor roles on real checkpoints: 98.7% (macro) on 2,269 repos of 167 model types, labels carried over from the corpus; 92% of repos reach 99%. Name rules with MoE: 95.3%, llama.cpp table: 33.2%.
87% of popular repos with no safetensors or PyTorch weights are GGUF-only (llama.cpp format).
Quantized copies are 6% of Hub repos but take 28% of downloads (×4.47); originals 71% of repos, 55% of downloads.
Derivatives keep the architecture: 73% of fine-tunes and 84% of quantized copies have the parent's width, depth, heads, FFN and vocabulary; 21% of fine-tunes change the vocabulary.
Lineage: 23% of parent links among popular repos point to Qwen models.
KV cache per token per billion parameters fell 24.3× from <=2022 to 2026; models whose 32k-token cache outweighs their 4-bit weights: 93% → 34%.
Experts multiply, active share falls: median experts 8 → 256, top-k 2 → 8, active share of parameters 28% → 11% (2024 → 2026).
How released models scale: width ∝ N^0.403, depth ∝ N^0.275 (dense LMs, n=965): larger models are wider rather than proportionally deeper.
13% of popular models above 1B parameters still store FP32: 9.78 TB that BF16 would halve.
Fits a 16 GB laptop at 4-bit, 8k context: 96% of original LMs released <=2022, 62% of 2026 releases.
Config fields agree with the weight shapes in 97% of checked repos; the Hub's parameter count is wrong for 12% and missing for 24%; 12% of popular repos declare no license.

Every number above, with its 95% interval and trend test, is in insights/ecosystem.json (written by ecosystem.py): Wilson intervals for shares, bootstrap intervals for medians, Cochran-Armitage and logistic trends over release year, Spearman for monotone trends, log-log slopes for shape scaling, Mann-Whitney for two groups.

What it can be used for

Architecture trends with evidence from the weights themselves: mixture of experts, hybrid state-space stacks, grouped-query and latent attention, head size, vocabulary, by release year and provider.
Metadata quality: repos whose declared config disagrees with the weights, and formats where the Hub's own parameter count is not the model's parameter count.
Deployment planning at scale: stored precision, KV-cache bytes per token and parameter counts for thousands of checkpoints without downloading one.
Benchmarking architecture-recovery and model-search tools against config_hp and family_config.
A snapshot to compare against: re-run the census later (the code and the frozen rules are public) to see what changed.

Summary tables

Repos probed: 12974 · crawled 2026-11-03T03:58:40Z → 2026-11-03T07:18:21Z

What happened when we read each repo

status	Top 10000 by downloads	Hub-wide (weighted, n=2974)
ok	67.3% ± 0.9	48.0% ± 2.0
gated	1.6% ± 0.2	1.0% ± 0.4
no_weights	27.0% ± 0.9	49.0% ± 2.0
legacy_pickle	2.9% ± 0.3	1.2% ± 0.4
not_found	0.0% ± 0.0	0.1% ± 0.1
timeout	0.8% ± 0.2	0.1% ± 0.1
network	0.0% ± 0.0	0.0% ± 0.0
error	0.4% ± 0.1	0.6% ± 0.3

Among repos we could read

	Top 10000 by downloads	Hub-wide (weighted, n=2974)
has a config.json	85.2% ± 0.8	72.2% ± 2.5
no config.json (weights are the only spec)	14.8% ± 0.8	27.8% ± 2.5
config disagrees with weights (≥1 checked field)	1.9% ± 0.3	1.4% ± 0.7
quantized storage	15.2% ± 0.9	5.3% ± 1.3
mixture-of-experts	7.7% ± 0.6	2.4% ± 0.9
has SSM / linear-attention layers	11.3% ± 0.8	3.8% ± 1.1
manifest family = config family	96.1% ± 0.5	96.3% ± 1.1
param count = Hub's count (±0.1%)	66.0% ± 1.1	54.1% ± 2.8

Cost of reading

header bytes read, median: 75 KB (p95 1.3 MB)
fraction of the repo's weight bytes read, median: 0.0026%
seconds per repo, median: 1.8s (p95 20.9s)
time spent waiting on rate limits: 0s total

Family by weights alone (Hub-wide, weighted share, 95% CI, n = repos behind it)

family	share	n
decoder	35.2% ± 2.7	574
encoder	26.7% ± 2.5	396
prefix-lm	12.7% ± 1.9	238
unknown	14.7% ± 2.0	189
encoder-decoder	9.3% ± 1.6	125
diffusion	1.3% ± 0.6	51
transducer	0.2% ± 0.2	3

Popularity vs readability

status	median downloads (30d)	median likes	repos
error	13,918	4	55
gated	15,534	85	227
legacy_pickle	25,294	14	321
no_weights	11,436	5	3968
not_found	3,950	0	4
ok	18,291	11	8302
timeout	17,838	92	97

Our parameter count vs the Hub's

reason	repos	median ours / Hub
equal	5363	1.000
other (check)	17	0.924
quantized: Hub counts packed storage elements	225	1.802
quantized: we undercount (check)	104	0.974
several components or copies (check)	348	1.393
no Hub count at all	2223	-

Why there were no weights to read

kind	Top 10000 by downloads	Hub-wide (weighted, n=2974)
gguf_only	86.3% ± 1.3	12.6% ± 1.9
no_weight_files	4.3% ± 0.8	49.2% ± 2.8
adapter	0.7% ± 0.3	28.8% ± 2.5
onnx_only	6.9% ± 1.0	1.3% ± 0.6
keras_or_other	1.4% ± 0.4	6.3% ± 1.4
other	0.4% ± 0.2	1.8% ± 0.7

Original releases vs derivatives (repos read in full, unweighted)

origin	repos	readable	no config.json	family by weights = family by config
adapter	213	70%	83%	100%
finetune	1354	91%	18%	97%
merge	65	86%	25%	98%
original	5233	84%	11%	96%
quantized	3135	29%	15%	96%

Weights vs config.json, field by field (repos with a config)

Each cell: weights alone / with config fused (heads, kv_heads, head_dim need the config to split width into heads × size). CI fixtures (tiny random test repos) and placeholder values (0) are excluded.

field	Original releases (top) (n=3506)	Frontier providers, recent (n=494)	Created in 2026 (n=1751)
d	97.3% / 97.3% (2853)	98.1% / 98.1% (431)	96.1% / 96.1% (1530)
layers	94.6% / 94.6% (2820)	88.9% / 88.9% (422)	92.1% / 92.1% (1514)
heads	81.0% / 97.3% (2799)	75.2% / 97.4% (427)	78.9% / 96.1% (1520)
kv_heads	75.9% / 98.8% (1053)	69.2% / 98.4% (380)	74.9% / 98.1% (1253)
head_dim	80.4% / 98.0% (496)	76.1% / 98.2% (285)	82.3% / 97.7% (1049)
ffn	95.7% / 95.7% (2351)	95.6% / 95.6% (251)	92.3% / 92.3% (1063)
vocab	93.0% / 93.0% (2633)	92.6% / 92.6% (408)	92.3% / 92.3% (1449)
experts	93.8% / 93.8% (210)	97.0% / 97.0% (164)	91.8% / 91.8% (401)
family	95.9% (3506)	92.9% (494)	94.9% (1751)

Most common failure messages

`` × 3931
worker exited -11 × 39
SystemError: <method 'eval' of '_quickjs.Context' objects> returned a result wit × 7
RuntimeError: unreadable index file × 3
RuntimeError: compressed data.pkl is not supported × 2
RuntimeError: not a safetensors file × 2
google/diffusiongemma-26B-A4B-it: no .safetensors or PyTorch checkpoint found × 1
peft-internal-testing/tiny-OPTForCausalLM-lora: no .safetensors or PyTorch check × 1
google/ddpm-cifar10-32: no .safetensors or PyTorch checkpoint found × 1
Manjushri/pornworks-real-porn-photo-realistic-nsfw-sdxl-and-pony-chekpoint: no . × 1

Citation

If you use this dataset, please cite:

@misc{singkul2026reading,
  title  = {Reading the Architecture off the Weights: Manifest-Only Model Anatomy and a Census of 12,974 Hugging Face Repositories},
  author = {Singkul, Sattaya},
  year   = {2026},
  note   = {Preprint},
  url    = {https://www.researchgate.net/publication/417293353}
}

S. Singkul, Reading the Architecture off the Weights: Manifest-Only Model Anatomy and a Census of 12,974 Hugging Face Repositories, 2026. PDF · ResearchGate · code: https://github.com/JoesSattes/ai-model-explorer

Downloads last month: 136

Space using JoesSattes/hub-weight-census 1

Collection including JoesSattes/hub-weight-census

AI Model Explorer

Collection

Model architecture read from safetensors headers: the labelled corpus, the Hub Weight Census and the explorer. • 3 items • Updated 3 days ago