Open-source large language models (LLMs) are AI models that use publicly available natural language text and software program code data to learn, understand and replicate human language.
Open Source LLMs, a form of open-source AI, can be freely modified and shared, enabling collaboration and customisation without licensing fees. This openness promotes transparency, safety, competition and diverse applications across industries. A form of open-source AI, these LLMs can be modified and shared freely, allowing collaboration and customisation for any purpose without permission or licensing fees — promoting transparency and safety, competition and polyculture and diverse applications.
Large language models are the foundation of modern user interfaces powered by artificial intelligence. These models, trained with massive datasets and advanced neural network architectures, make it possible for humans to interact with applications on a natural level. The value of LLMs lies in their ability to perform nuanced language tasks, such as summarisation of large passages of text or, conversely, generating large bodies of text in response to their human user's instructions. In fields as diverse as customer service, research, content creation and education, LLMs allow users to make complex queries and instructions, and for the AI to respond in a way that even inexperienced users can comprehend at super-user equivalency in some cases.
But while these models may allow for user-friendly interactions, they are extremely complex, often built on proprietary data that reduces visibility into their internal structure and capabilities. While open-source LLMs generally speaking offer a more accessible alternative, the degrees of openness vary. While some large language models throw back the curtain on the underlying training code and data, allowing anyone to use, modify or distribute it, others may only release the model and limited supporting artefacts, making auditing the upstream model supply chain near impossible. By democratising access to powerful AI tools, open-source LLMs give a broader range of developers the power to innovate and customise AI solutions to their specific needs.
Any discussion of open-source LLMs begs the question: How are open-source LLMs different from proprietary LLMs? Open-source LLMs and proprietary LLMs differ significantly in transparency, accessibility, adaptability and community engagement.
Open-source LLMs are freely available to be used and modified as any user sees fit and may be shared (modifications and all) with others. This encourages a collaborative approach to downstream innovation, allowing developers to customise these models to meet specific needs while also contributing to a dynamic ecosystem where improvements and new applications are continuously emerging. Essentially, every developer that improves on the base model and open sources their fine-tuned model becomes a member of the AI ecosystem for that community.
For this to be possible, open-source LLMs rely heavily on transparency in the model's architecture, training and intended use which it was designed to support. To maximise community and commercial adoption, the data used for pre-training and evaluation, the resources involved and the underlying code itself must be fully accessible for review. This is a major differentiating factor in open-source LLMs, as proprietary LLMs typically lack visibility into their inner workings.
Open-source LLMs offer enhanced freedom in the form of flexibility; organisations can tailor the LLM to their unique specifications. On the other hand, changes introduced in the derivative LLM could result in weakened security, so it is important to select open-source LLMs that have robust data and model governance practices in place to ensure new models meet safety and performance expectations for downstream users. Working with an open-source solution can involve significant costs, including hiring and training expertise, upfront legal fees, feature upgrades, security compliance, support, talent retention and ongoing software life cycle management.
Proprietary LLMs are controlled and owned by individual entities, with access typically restricted through licences and fees. Companies like OpenAI and Google offer powerful LLMs, but their use is often limited to predefined application programming interfaces (APIs) or specific applications dictated by the providers. This closed approach can limit customisation and adaptation, potentially increasing costs and limiting access to inference cloud computer infrastructure that may not be in the same region as the end user. Additionally, proprietary LLMs may build on fine-tuned versions of open-source models, with companies adding unique enhancements or 'secret sauce'—such as improved performance or specialised functionalities—that make these versions proprietary. An example of this approach includes offerings like Now LLMs, where tailored improvements set them apart from open-source alternatives.
That said, there are some advantages to working with proprietary LLM solutions. Specifically, owned and licensed LLMs tend to offer stronger security and are more user-friendly, with company-backed support providing assistance when needed.
Open-source software is nothing new; in fact, some of the earliest software was shared freely between those who worked on computers. But with the relatively recent advent of artificial intelligence, the concepts behind open source are beginning to take root in AI. Still, although they share core philosophies of transparency, collaboration and accessibility, there are also some significant differences:
Open-source software refers to programs or systems whose source code is made freely available for anyone to use, modify and share. Open-source software is governed by licences like the GNU General Public License (GPL) or Apache License, which outline the terms of use and redistribution. Popular examples include Linux, Apache HTTP Server and Mozilla Firefox, which continue to thrive through contributions from global developer communities.
With the rise of artificial intelligence, the principles behind open-source software have been adapted to AI. Open-source AI is the result. The Open Source Initiative defines open-source AI as: "An AI system made available under terms and in a way that grant the freedoms to:
These freedoms apply both to a fully functional system and to discrete elements of a system. A precondition to exercising these freedoms is to have access to the preferred form to make modifications to the system."
- Use the system for any purpose and without having to ask for permission.
- Study how the system works and inspect its components.
- Modify the system for any purpose, including to change its output.
- Share the system for others to use with or without modifications, for any purpose.
These freedoms apply both to a fully functional system and to discrete elements of a system. A precondition to exercising these freedoms is to have access to the preferred form to make modifications to the system."
Like open-source software, open-source AI promotes transparency and collaboration by allowing developers direct access to the AI's code. Open-source licences play a central role in this ecosystem. Licences (such as the Blue Oak Model License) outline the terms and conditions associated with using the AI. These legal frameworks help keep open-source AI accessible while protecting contributors from liability.
Open-source LLMs employ open-source AI supported by licences to provide developers with the freedom to customise and adapt large language models to their specific needs without the constraints of proprietary systems.
There are many options for organisations that are interested in working with open-source LLMs. The following are some of the most transparently developed open-source language models currently available:
StarCoder, developed by the BigCode project open-scientific collaboration led by Hugging Face and ServiceNow, is an open-source LLM designed for code generation. Trained on over 80 programming languages, it excels in code generation, workflow generation and even text summarisation tasks. StarCoder has a large context window and unique features like infilling capabilities, making it a strong choice of foundation model. It is licensed under OpenRAIL-M, which allows for free commercial use with ethical and responsible AI use case restrictions.
Created by the German AI startup Aleph Alpha, Luminous focuses on delivering advanced natural language understanding and generation capabilities. It is designed to compete with advanced LLMs (such as recent versions of ChatGPT) while offering transparency and ethical AI development. Luminous consists of 13 billion parameters and is available for tasks ranging from small to large-scale language applications.
IBM's Granite models are open-source code LLMs designed for enterprise-grade applications. Granite models are trained on 116 programming languages and can be used for code generation and bug fixing along with more traditional summarisation and explanation. They are released under the Apache 2.0 license, making them suitable for both research and commercial use.
Phi-2, developed by Microsoft Research, is a lightweight, 2.7 billion-parameter model that nonetheless is capable of competing with many larger models in terms of performance. Phi-2 can perform complex tasks, including commonsense reasoning, coding, maths and language understanding. Its compact size makes it ideal for efficient experimentation and it is available in the Azure AI Studio.
Building on the success of the Stable Diffusion framework, Stable Video Diffusion is an open-source model focused on video generation and editing. It leverages AI to create dynamic visual content, offering powerful tools for industries like entertainment and advertising.
Meta's Llama 3 (with 3.1 being the most recent release) is the latest in the Llama family of models. Available in model sizes ranging from 8 billion to 405 billion-parameter versions, Llama 3 builds on the architecture of its predecessors. Llama 3 performs well in reasoning, coding and multilingual tasks, and has expanded safety tools or detecting insecure code
Developed by Google, BERT (Bidirectional Encoder Representations from Transformers) is an encoder-only transformer model designed to understand and generate human language. Since its release in 2018, BERT has been widely adopted for various natural language processing (NLP) tasks such as text classification, question answering and sentiment analysis. Despite its relative age, its influence continues to shape modern NLP.
Versatile and accessible, open-source LLMs have a wide array of applications across a range of industries. Here are some prominent ways in which they can be applied:
Open-source LLMs can be used to effortlessly transform PDFs into engaging podcast-style audio. By implementing a compound LLM workflow, you can convert text from PDFs into a conversational script and then narrate it using advanced text-to-speech technology. This is ideal for creating accessible content, educational materials or just bringing your documents to life in a new format.
It can be hard to gauge the emotion behind customer feedback. Open-source LLMs can be trained to analyse text to determine the overall sentiment being expressed: positive, negative or neutral. This allows businesses to better understand customer responses to improve their products and services. Sentiment analysis is a major tool in monitoring social media for customer insights.
Many open-source LLM models can assist developers by providing code suggestions, writing complex algorithms, fixing bugs in code and even documenting code. They can generate code snippets from natural language descriptions, allowing users to tell the code what they want it to do in plain English.
LLMs are widely used for generating coherent and contextually relevant text. This includes creating articles, stories and dialogue for virtual assistants. This is a major part of Generative AI (GenAI), and has applications in essentially every arena.
Open-source LLMs can automate content creation and generate summaries of lengthy documents, helping users to quickly grasp the main points of extensive texts. This is particularly valuable for professionals who need to process large volumes of information efficiently.
LLMs are helping break down language barriers. With their multilingual capabilities, many open-source LLMs can translate text between multiple languages, making clear communication with a global audience possible. These models are trained on diverse language datasets, ensuring accurate and context-aware translations that go well beyond simple word-by-word translations.
Open-source LLMs power today's AI chatbots and virtual assistants, enhancing customer support by providing quick and accurate responses to queries. When designed for conversational tasks, LLMs can handle customer interactions effectively and in a way that feels natural for the user. And should there be anything the LLM is not prepared to address, it can be automated to escalate the issue to a human agent.
LLMs can be integrated into educational platforms to provide personalised learning experiences. They can adapt content to individual learning styles, offer explanations and generate practice problems specifically tailored to the needs and capabilities of individual users.
- Healthcare
LLM-powered AI telemedicine solutions provide always-available virtual caregivers capable of diagnosing, providing information and organising patient information. Much like AI chatbots and virtual agents, these programs are designed to evaluate patient issues based on insights from extensive data sets, so they know when to provide assistance on their own and when to alert human medical teams.
- Finance
In the financial sector, open-source LLMs enhance fraud detection, automate customer support and perform sentiment analysis to identify emerging trends. These models analyse financial documents and other data to provide real-time market insights.
- Journalism and news
Journalists and news organisations utilise open-source LLMs for summarisation and translation. LLMs can be used internally to analyse information without sharing proprietary data outside the newsroom. And, for when circumstances demand an immediate turn-around, news agencies can direct open-source LLMs to generate relevant and informative content that speaks to their intended audience.
- Science-based industries
LLMs support scientific research by automating literature reviews, data analysis and hypothesis generation, but they can also do so much more. The extreme flexibility of these models means scientists can adapt them to any kind of research — from fighting climate change to analysing DNA sequences and modelling astrophysical phenomena.
Customisable autonomous systems that can follow human-language directions and respond in kind — it is not hard to imagine what kinds of advantages these models can mean for business. Companies that leverage open-source LLMs can expect to see:
While open-source LLMs offer many benefits, they can also present certain risks. Understanding the following challenges and how to counter them is an important part of ensuring responsible and effective use:
Open-source LLMs, due to their unrestricted access, can be used for harmful purposes as easily as beneficial ones. Spreading misinformation, violating privacy, accessing restricted or proprietary information — these are all ways in which an AI can be exploited. Ensuring ethical use requires proactive community governance and clear guidelines to balance innovation with safety and responsibility.
Community contributions to open-source LLMs can vary in quality, leading to inconsistent outputs. Without standardised oversight, these models might yield unreliable results, particularly in applications where accuracy is critical. Maintaining high standards means keeping a close eye on all code contributions and refining them when necessary. Unfortunately, this can be a major job when working with open-source codes.
Any AI is only as good as the information it's built on, and when that information includes prejudiced opinions or unfair assumptions, the AI may learn to perpetuate them. LLMs can become influenced by the biases present in their training data, resulting in skewed and unfair outputs. Vigilant oversight and diverse data curation are essential to mitigate bias and ensure fairness and inclusivity in AI-generated content and decisions.
When selecting an open-source large language model (LLM), organisations need to consider several factors to ensure they choose the best fit for their needs. Here are the most important considerations:
- Goals
Identify the main purpose of the LLM. What is it going to achieve for the business? What will its focus be? Some models are tailored for research, while others are suited for commercial use. Make sure the LLM aligns with specific goals.
- Accuracy
Evaluate the model's accuracy for the tasks it will face. Even with access to customise the code, some LLMs are better suited for different uses.
- Cost
While open-source LLMs are free, consider the costs of hosting, training and maintaining the model. Larger models require more resources, which can increase infrastructure and operational expenses.
- Performance
Assess the LLM's language fluency, coherence and context comprehension. High performance in these areas improves user experience and task effectiveness.
- Data Security
Ensure the LLM can securely handle sensitive data, especially when dealing with personal or proprietary information.
- Quality of training data
Check the quality of the original training data used by the LLM. High-quality data leads to better outputs. If necessary, use custom data to fine-tune the model for improved results.
- Available skillsets
Complex LLMs require advanced skills in data science, machine learning operations (MLOps) and NLP. Ensure the team that will be working most closely with the LLM has the necessary experience. If they do not, make plans to train or hire to fill that skill gap.
ServiceNow delivers top-tier proprietary language models through the Now LLM Service, enabling advanced AI-driven features like chat summarisation, agent record summarisation, AI-enhanced search, dynamic translation and seamless conversational flows. These capabilities are built on open-source innovation; ServiceNow openly trains foundation models, such as StarCoder, as part of its initiatives. These models are then refined into proprietary Now LLMs, specifically tailored to optimise productivity across the enterprise and deliver exceptional generative AI experiences.
This open-source foundation, enhanced through collaboration with partners like Hugging Face and NVIDIA, allows ServiceNow to advance AI responsibly while sharing valuable innovations with the community. Through this approach, organisations benefit from powerful, specialised AI skills — from automated workflow recommendations to precise text-to-code generation. And, because Now LLMs are part of the fully integrated Now Platform®, they seamlessly enhance existing workflows and applications to drive efficiency across every department.
See ServiceNow in action and discover how AI and LLM can enhance your operations; schedule your demo today!