Updated 2026-09-20 · ranked by real download data

Models that fit in 12 GB

Ranked nightly from our download snapshots of the Hugging Face catalog. Every entry shows its licence and what hardware it realistically needs.

#	Model	Params	Context	Commercial use	30d	Min VRAM
01	all-MiniLM-L6-v2 +1 variant sentence-transformers · sentence-similarity	20 M	512	✓ apache-2.0	254.1 M	from 0.6 GB
02	bge-small-en-v1.5 BAAI · feature-extraction	30 M	512	✓ mit	64.5 M	from 0.7 GB
03	paraphrase-multilingual-MiniLM-L12-v2 sentence-transformers · sentence-similarity	120 M	512	✓ apache-2.0	45.9 M	from 1.0 GB
04	t5-small google-t5 · translation	60 M	—	✓ apache-2.0	24.9 M	from 0.8 GB
05	Qwen3-0.6B Qwen · text-generation	750 M	41 K	✓ apache-2.0	23.0 M	from 2.3 GB
06	all-mpnet-base-v2 sentence-transformers · sentence-similarity	110 M	512	✓ apache-2.0	22.6 M	from 1.0 GB
07	mobilenetv3_small_100.lamb_in1k timm · image-classification	<0.1 M	—	✓ apache-2.0	18.0 M	from 0.5 GB
08	bge-reranker-v2-m3 BAAI · text-classification	570 M	8 K	✓ apache-2.0	17.6 M	from 3.1 GB
09	gpt2 openai-community · text-generation	140 M	—	✓ mit	15.4 M	from 1.1 GB
10	nomic-embed-text-v1.5 nomic-ai · sentence-similarity	140 M	2 K	✓ apache-2.0	14.7 M	from 1.1 GB
11	efficientnet_b3.ra2_in1k timm · image-classification	10 M	—	✓ apache-2.0	12.7 M	from 0.6 GB
12	multilingual-e5-small intfloat · sentence-similarity	120 M	512	✓ mit	12.3 M	from 1.0 GB
13	Kokoro-82M hexgrad · text-to-speech	82 M	—	✓ apache-2.0	11.6 M	from 0.7 GB
14	bge-large-en-v1.5 BAAI · feature-extraction	340 M	512	✓ mit	11.4 M	from 2.0 GB
15	bge-base-en-v1.5 +1 variant BAAI · feature-extraction	110 M	512	✓ mit	10.5 M	from 1.0 GB
16	paraphrase-multilingual-mpnet-base-v2 sentence-transformers · sentence-similarity	280 M	512	✓ apache-2.0	9.9 M	from 1.8 GB
17	Qwen3-Embedding-0.6B Qwen · feature-extraction	600 M	33 K	✓ apache-2.0	8.6 M	from 1.9 GB
18	Qwen2.5-0.5B-Instruct Qwen · text-generation	490 M	33 K	✓ apache-2.0	8.5 M	from 1.7 GB
19	clap-htsat-fused laion · audio-classification	150 M	—	✓ apache-2.0	8.2 M	from 1.2 GB
20	multilingual-e5-base intfloat · sentence-similarity	280 M	512	✓ mit	7.6 M	from 1.8 GB
21	opt-125m facebook · text-generation	125 M	2 K	other	7.4 M	from 0.8 GB
22	Qwen2.5-1.5B-Instruct Qwen · text-generation	1.5 B	33 K	✓ apache-2.0	7.2 M	from 4.1 GB
23	vit-base-patch16-224 google · image-classification	90 M	—	✓ apache-2.0	7.2 M	from 0.9 GB
24	multilingual-e5-large intfloat · feature-extraction	560 M	512	✓ mit	7.2 M	from 3.0 GB
25	Qwen3-4B Qwen · text-generation	4.0 B	41 K	✓ apache-2.0	7.1 M	from 10.0 GB

Membership and ranking refresh nightly after the snapshot run. VRAM is an estimate for the smallest quantization at 8K context — methodology.