Overview

Transformers 4.48, released on January 15, 2025, adds DeepSeek model support and improves the text generation pipeline.

Main Features

DeepSeek support

DeepSeek models are natively integrated into the library, allowing direct usage via the AutoModel API.

python 
from transformers import AutoTokenizer, AutoModelForCausalLM

tokenizer = AutoTokenizer.from_pretrained('deepseek-ai/deepseek-coder-1.3b-base')
model = AutoModelForCausalLM.from_pretrained('deepseek-ai/deepseek-coder-1.3b-base')

inputs = tokenizer('def fibonacci(n):', return_tensors='pt')
outputs = model.generate(**inputs, max_new_tokens=50)
print(tokenizer.decode(outputs[0]))

Improved generation

The generate() pipeline is optimized with better batching support and decoding strategies.

python 
from transformers import pipeline

gen = pipeline('text-generation', model='gpt2')
result = gen(
    'Python is',
    max_new_tokens=30,
    num_return_sequences=2,
    do_sample=True,
    temperature=0.7,
)
for r in result:
    print(r['generated_text'])

Sources