Comparing and contrasting the 3 most popular Open Source LLMs: BERT, Bloom, and Vicuna

3 min readJun 11


In the ever-evolving field of natural language processing (NLP), BERT, Bloom, and Vicuna have emerged as powerful models, each with its own unique characteristics.

Understanding the differences in their training methods, architectures, pre-training tasks, strengths, and use cases is vital for selecting the most suitable model for specific NLP applications.

In this article, we will delve into the details of BERT, Bloom, and Vicuna, shedding light on their key features and exploring the breadth of their applications.

We know this can be a complex topic so if you would like a free 25 minute consultation on which Open Source LLM is best for your needs, book a call with us at


Training Method

BERT (Bidirectional Encoder Representations from Transformers) employs unsupervised learning and pre-training to create contextualized word representations.

It learns from large amounts of unlabeled text data by predicting masked words within sentences and utilizing the transformer architecture.


BERT employs the transformer architecture, which consists of multiple layers of self-attention and feed-forward neural networks. This architecture allows for efficient capturing of contextual dependencies in language.

Pre-training Task

The pre-training task for BERT involves two main objectives: masked language modeling and next sentence prediction.

Masked language modeling requires the model to predict masked words in a sentence, while next sentence prediction involves determining whether two sentences follow each other in the original text.

Strengths of BERT

  • Contextual understanding of language
  • Transfer learning capabilities
  • Effective for a wide range of NLP tasks
  • Large pre-trained models available

Limitations of BERT

  • High computational requirements and memory consumption
  • Inability to handle long documents efficiently
  • Lack of interpretability for the learned representations

2. Bloom

Training Method

Bloom filters are not trained models; they are probabilistic data structures designed to test whether an element belongs to a set. The filter is constructed using a series of hash functions and bit arrays.


Bloom filters have a simple architecture based on hash functions and bit arrays, which allows for efficient membership testing.

Pre-training Task

N/A (Bloom filters are not trained models)

Strengths of Bloom

  • Compact memory footprint
  • Fast lookup times
  • Scalability for large datasets
  • Minimal false negatives

Limitations of Bloom

  • Not suitable for fuzzy matching or complex language tasks
  • Limited functionality beyond set membership testing

3. Vicuna

Training Method

Vicuna utilizes unsupervised pre-training and transfer learning to excel in multilingual machine translation. It leverages the successes of models like BERT to achieve accurate translations across multiple languages.


Vicuna’s architecture builds upon the advancements made in transformer models, similar to BERT. It focuses on multilingual translation tasks.

Pre-training Task

N/A (Vicuna is trained specifically for multilingual machine translation)


  • Multilingual translation capabilities
  • Efficient cross-lingual transfer learning
  • Improved translation accuracy and fluency
  • Ability to handle low-resource languages


  • Multilingual machine translation
  • Cross-lingual information retrieval
  • Language localization


BERT, Bloom, and Vicuna exemplify the diversity and ingenuity present in the world of NLP. BERT’s contextual understanding and transfer learning have propelled it to the forefront of NLP research and applications.

Bloom’s compact design and efficient membership testing make it an ideal choice for information retrieval and set membership tasks. Vicuna, with its focus on multilingual translation, addresses the challenges of language diversity in a connected world.

Understanding the training methods, architectures, pre-training tasks, strengths, and use cases of these models is crucial for selecting the right tool for specific NLP requirements.

Once again, we know this can be a complex topic so if you would like a free 25 minute consultation on which Open Source LLM is best for your needs, book a call with us at




We build custom, secure, robust chat bots for security & privacy minded enterprises