Machine Learning & Large Language Models

Large Language Models (LLMs)

Over the past few years, Large Language Models have become a very important discussion point since the release of ChatGPT to the public in late 2022. Since the Release of 2022's ChatGPT based on GPT 3.5, LLMs have come a long way, with ChatGPT itself moving to improved models such as 4o, and open source solutions being introduced. While it is unfortunate that many programs such as ChatGPT are based on closed source models, in recent months there has been a surge of interest regarding new Open Source Models.

 

Finding Open Source LLMs

As interest in Large Language Models continues to grow, there are now hundreds of models available for public use. However, not all “open” models are truly open source, and understanding the differences is important when selecting a model for your project.

One of the best places to explore available LLMs is Hugging Face, a widely used platform that hosts thousands of models. You can filter models by:

  • License (e.g., MIT, Apache 2.0, custom licenses)
  • Task (text generation, summarization, etc.)
  • Model size and performance

Sorting by license is especially important if you plan to use a model for research, personal projects, or commercial applications.

Not all LLMs labeled as “open” provide the same level of access. There are three main categories and understanding their distinctions helps ensure you are using a model that aligns with your goals and any licensing requirements.:

  • Fully Open Source Models
    These models provide access to both the model weights and training code, often under permissive licenses like MIT or Apache 2.0.
    → Example use: research, modification, and redistribution
  • Open-Weight Models
    These models (like Meta’s Llama series) provide access to model weights but have restrictions on usage, redistribution, or commercial deployment.
    → Common in industry-backed models
  • Closed Source Models
    These models (like ChatGPT) do not provide access to weights or training data and are only accessible via APIs.

Some widely used models you may encounter include:

  • Llama (Meta) – Highly capable, widely adopted, but with a custom license
  • Mistral – Efficient and performant open models
  • Gemma (Google) – Lightweight and research-friendly
  • Falcon – Strong performance with open licensing options

Each model varies in size, performance, and hardware requirements, so choosing the right one depends on your specific use case.