Less Risk and No Fee: Overview of Open-Source LLMs for Enterprises

What do VMware, IBM, Walmart, and Shopify have in common? Apart from their considerable global presence, these enterprises all use open-source large language models.

And they aren't anomalies in a world dominated by proprietary software. Experts increasingly believe that open-source GenAI can outperform its closed-source counterparts. One such expert is a software engineer at Google.

In a leaked document that gained widespread attention last year, he argued that open-source models are on track to outcompete those from tech giants like Google and OpenAI. “While our models still hold a slight edge in terms of quality, the gap is closing astonishingly quickly. Open-source models are faster, more customizable, more private, and pound-for-pound more capable. They are doing things with $100 and 13B params that we struggle with at $10M and 540B.”

One year later, proprietary GenAI still leads the way. But the open-source community hasn't been idle either: Faster, more accurate, and more efficient large language models (LLMs) — which are available for free — have come onto the market.

So maybe you don't need to invest in yet another GenAI solution that puts your enterprise data and reputation at risk without delivering improved accuracy. For your convenience, we’ve compiled a list of state-of-the-art open-source LLMs with use cases for enterprises.

But first, let’s revisit the basics of open-source LLMs.

What Are Open-Source Large Language Models and Their Main Benefits?

Open-source LLMs are developed and released with publicly available source code. This makes them different from proprietary GenAI, whose internal workings remain hidden from the public.

Open-source large language models appeal to businesses for many reasons:

• Transparency. Users can view the code to understand how the model works, which mitigates regulatory concerns and promotes trust.
• Privacy and control. Enterprises can deploy and fine-tune models on their own infrastructure and ensure that their confidential information remains secure and private.
• Adaptability. Open-source LLMs are easier to customize for specific use cases than proprietary models, enabling better output quality.
• Cost efficiency. These models are free to use, which reduces the cost associated with adopting AI.
• No vendor lock-in. When using open-source models, you’re not tied to long-term contracts with proprietary vendors — you can switch vendors flexibly.

While open-source LLMs are not always known for their performance and high quality of outputs, their customization potential can mitigate this issue. According to venture capital firm Andreessen Horowitz, one company found that “after fine-tuning, Mistral and Llama perform almost as well as OpenAI but at a much lower cost.”

The Current State of Open-Source GenAI

Due to the challenges of commercializing open-source software, closed-source GenAI dominated the market in 2023, with OpenAI’s models boasting an 80% share globally. However, in 2024, the landscape is shifting as companies increasingly turn to open-source GenAI.

Numerous studies highlight this shift. Andreessen Horowitz surveyed over seventy top enterprise leaders to understand their strategies for GenAI adoption. The findings revealed that:

• 46% of respondents preferred or strongly preferred open-source LLMs.
• Looking ahead to 2024 and beyond, the surveyed enterprise leaders anticipated a noticeable shift toward open-source GenAI, with some aiming for a 50/50 split between closed-source and open-source solutions, up from an 80/20 split in 2023.

The increasing demand for open-source GenAI powers up the supply of open-source models. In 2023, open-source AI startups received $2.9 billion in funding in venture capital — an increase from $900 million in 2022. Tech giants are becoming players in this space, with Meta, Google, and other big names launching their own open-source LLMs.

The trend is clear: more open-source GenAI solutions are entering the market. Let’s explore what’s already available.

Top Six Open-Source LLMs

Whether you want to develop a GenAI product for customers or an internal solution, the open-source market offers a variety of options. The following list highlights examples of the best open-source LLM solutions for the most common GenAI use cases in enterprises.

Llama 3 

Llama 3 is a decoder-only transformer LLM released by Meta in April 2024. It is the latest version of the Llama family. The LLM has been trained on Meta’s custom-built 24K GPU clusters on more than 15 trillion tokens of text (over 30 languages) and code data and has a context length of over 8,000 tokens. It is available in 8B and 70B versions.

Llama 3 is designed as a multimodal LLM and can be customized for a variety of text and code generation tasks. In addition, image and animation functions will reportedly be available soon. The model outperforms Gemini Pro 1.5 (Google’s proprietary model) on various benchmarks across use cases, including asking for advice, brainstorming, classification, answering open and closed questions, coding, creative writing, extraction, reasoning, rewriting, summarising, and more. It also outperforms Mistral Medium, GPT-3.5, and Claude Sonnet in the majority of cases. 

Licensed under the “Meta Llama 3 Community License Agreement,” Llama 3 can be used for research and commercial purposes. Companies can use the LLM to create emails, product descriptions, web content, and marketing texts. It is also a perfect foundation for AI assistants, and the new Meta AI (running Llama 3) is a confirmation of this. However, Meta's approval is required to develop products with over 700 million monthly users.

Mistral 

Mistral AI is a French AI software company founded by former employees of Meta Platforms and Google DeepMind. Among Mistral’s three open-source LLMs, Mixtral 8x22B is the most powerful.

Based on SMoE (sparse mixture of experts) architecture, the model uses many expert subnetworks, each specializing in a specific area of data processing, but it only needs to activate a small subset of them for each output. That is, out of 141 billion parameters, the model may activate only 39 billion. Thanks to its sparse activation patterns, Mixtral 8x22B's cost efficiency for its size is unmatched among open-source models and performance-to-cost ratio. It is also faster than any dense 70B model and more powerful than any other open-weight model.

Mixtral 8x22B excels in translation and content creation in English, French, Italian, German, and Spanish. Its context length of 64,000 tokens makes it particularly valuable for summarizing large documents, generating stories, and answering long questions. It also shows strong capabilities in math and function calls.

The model outperforms Llama 2 for all the capabilities mentioned above on external benchmarks. Although official benchmarks are not yet available, the preliminary evaluation suggests that Mixtral 8x22B can match the performance of OpenAI’s GPT-4.

Gemma 

Gemma is a lightweight, dense, decoder-only LLM released by Google in February 2024. With a default context window of 8192 tokens, Gemma comes in two variants: 2B and 7B. Sharing the same technology as Google's Gemini, Gemma models offer an excellent performance-to-size ratio and can run on Google Cloud and NVIDIA GPUs.

Trained on a vast corpus of text (mainly in English), code, and math data, Gemma LLMs outperform larger open-source models in key benchmarks across various capabilities, including answering questions, reasoning, math, and coding. They surpass models like Llama 2 (7B and 13B) on benchmarks such as MMLU, HellaSwag, and HumanEval.

Google promotes Gemma models as secure and reliable and guarantees that personal and other confidential data was filtered out of the training data sets. To help engineers develop powerful and responsible AI systems with these LLMs, Google offers the Responsible Generative AI Toolkit and toolchains for supervised fine-tuning of Gemma LLMs in all major frameworks. 

However, Gemma models have limitations. Like Llama 3, they aren't open-source in the traditional sense; they are open models. This means that while Google grants public access to their weights, it retains control over how the models are used. For example, while the models are available for commercial use, Google places restrictions on the number of users. 

Vicuna 13B 

Vicuna 13B is an open-source LLM for chatbots with 13 billion parameters developed by a team from UC Berkeley, Carnegie Mellon University, Stanford, and the University of California, San Diego. The model was fine-tuned using Meta’s Llama 13B, and the fine-tuning involved 70,000 user-shared conversations from ShareGPT.

Vicuna 13B performs impressively, achieving 90% of the accuracy of GPT-4 and Google Bard. It also outperformed both Llama 13B and Stanford Alpaca in over 90% of the cases evaluated. The model's ability to generate human-like text, primarily in English, makes it an excellent choice for tasks that require advanced NLP skills, such as content generation, summarization, and answering questions.

However, Vicuna 13B is only available for non-commercial use. It can be used internally within an organization for research needs and other non-commercial purposes.

BLOOM 

If you’re looking for something truly universal in terms of human language comprehension, consider BLOOM — the first-ever multilingual LLM.

The BigScience Large Open-science Open-access Multilingual Language Model (BLOOM) is an autoregressive LLM featuring a decoder-only transformer architecture. Launched in 2022, BLOOM was developed through a global collaborative effort involving over 1,200 developers from 39 countries, coordinated by BigScience in partnership with Hugging Face and the French NLP community.

Trained on the ROOT corpus to continue text from a prompt, BLOOM boasts 176 billion parameters and requires industrial-scale computational resources to run. The model can comprehend and generate text in 46 languages, from English and French to Indonesian and Catalan. It’s also well-versed in 13 programming languages. Still, it has varying proficiency in each of the languages.

The model is a decent solution for multilingual content generation, code generation and debugging, and linguistic analysis. However, it’s not designed to process sensitive information or for applications that require the utmost precision, such as medical diagnosis or legal decisions.

Falcon 

Falcon is a family of LLMs released by the Technology Innovation Institute of the United Arab Emirates. The institute has rolled out three open-source models: Falcon 40B, Falcon 180B (which outperformed Llama 2 and GPT-3.5 on various tasks), and Falcon 2.

The latest model, Falcon 2, is the most lightweight LLM in the family, with only 11 billion parameters. It’s a multilingual and multimodal AI model with vision-to-language capabilities. It outscores Meta Llama 3 8B and performs comparably to Google’s Gemma 7B and Llama 2 70B. The model currently relies on a dense architecture, and the developer plans to release an SMoE version to enhance the model’s capabilities.

Trained on a whopping 5.5 trillion tokens, the Falcon 2 LLM is proficient in ten languages. Its uses include natural language understanding, text generation, machine translation, code generation, and information retrieval. It can be a powerful foundation for chatbots, content creation solutions, translation tools, recommendation systems, and more.

Conclusion

With data security being one of the main barriers to GenAI adoption, open-source LLMs offer enterprises a secure and transparent way to leverage this technology. These models enable companies to keep their sensitive data in-house and provide full transparency regarding how the solution works. In addition, the market is rich with state-of-the-art offerings that meet different needs, whether you’re developing a chatbot for customers, a copilot for the development team, or an AI assistant for the finance department.

Developing a full-fledged product involves considerable effort. If your company lacks the internal resources to optimize and seamlessly integrate an open-source LLM into your system, Dynamiq can help you here. Our low-code platform lets you fine-tune and deploy an LLM quickly and efficiently. The entire process takes place on premises using your proprietary training datasets and ensures that you maintain full control over the customized model.

For more information about how Dynamiq can support your enterprise, please contact us.