How leaders were counted
HuggingFace publishes downloads for the last 30 days for each model. Data taken from May 2026, filtered by categories: text-generation, embeddings, speech, vision.
Top 10 by category
Text Generation
| Model | Downloads/30d | License | For what |
|---|---|---|---|
| meta-llama/Llama-3.1-8B-Instruct | 18M | Llama | General purpose |
| Qwen/Qwen2.5-7B-Instruct | 12M | Apache 2.0 | Multilingual, code |
| mistralai/Mistral-7B-Instruct-v0.3 | 9M | Apache 2.0 | Code, instructions |
| deepseek-ai/DeepSeek-V3 | 7M | MIT | Complex tasks |
Embeddings
| Model | Downloads/30d | Size | MTEB Score |
|---|---|---|---|
| sentence-transformers/all-MiniLM-L6-v2 | 45M | 22MB | 56.3 |
| BAAI/bge-m3 | 22M | 570MB | 66.2 |
| nomic-ai/nomic-embed-text-v1.5 | 8M | 137MB | 62.4 |
Speech (Whisper)
openai/whisper-large-v3 — 31M downloads in 30 days. Best open-source speech recognition including Russian language. WER 5.1% on CommonVoice Russian. Whisper.cpp allows running on CPU in real time.
Vision
openai/clip-vit-large-patch14 — 28M downloads, standard for image-text matching. google/vit-base-patch16-224 — 19M downloads, image classification.
How to load any model
from transformers import AutoModelForCausalLM, AutoTokenizer
import torch
model_name = "meta-llama/Llama-3.1-8B-Instruct"
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForCausalLM.from_pretrained(
model_name,
torch_dtype=torch.float16,
device_map="auto" # автоматически распределяет по GPU/CPU
)
FAQ
Is HuggingFace Pro needed to download models?
Most models are free. Llama 3 requires accepting the license agreement on the Meta website. Gemma 2 — similarly via Google. Registration is sufficient.
How to choose a model for Russian language?
Qwen 2.5 7B — the best choice for Russian among models up to 10B. Among larger models: Llama 3.1 70B and DeepSeek V3 support Russian at a good level.