The LLM Index
A list of large language models (LLMs), including open-source and commercial offerings, comparisons of each, and libraries for working with LLMs.
Large language models (LLMs) are powerful machine learning systems that for many use cases can now understand and compose text at a human level. They are currently the leading subcategory of Foundation Models, large models pretrained using unsupervised methods on enormous datasets that can be tuned to perform a range of tasks. Due to their capabilities, individuals as well as businesses are now regularly using LLMs. This index is a list of LLMs and their properties and functionality. For a recent "evolutionary tree", we recommend Figure 1 in this paper.
Note that LLMs are being developed and released at a frantic clip. While we'll try and keep this LLM list up-to-date, we may have missed some recent releases. Please contact
zxie[at]sapling.ai with any significant updates.
Many reading this will be most interested in which LLM will perform best for their use case. While this can depend on the evaluation method and things are changing rapidly, we recommend the following resources to help make that assessment:
- AlpacaEval Leaderboard
- Chatbot Arena (LMSYS Org) and "Leaderboard" tab on LMSYS
- Open LLM Leaderboard (Hugging Face)
Most software businesses are familiar with cloud service providers (CSPs) that provide scalable computing resources. With the growth of ChatGPT, new LLM cloud services have been launched from familiar incumbents as well as well-capitalized startups.
|LM||Initial Release||Developer||Instruct / RLHF||Reference|
Open Source LLMs
Assuming you have the ability to run models with billions of parameters, using an open source model is one way to ensure control of your systems and data. The open source LLM ecosystem is moving quickly, most notably after the release of Meta's LLaMA models. In parallel to the release of powerful models trained on large corpuses of data and instruct-finetuned by research groups, a community of developers has also made it possible to run larger and larger models in real-time on commodity hardware—even, for example, on a consumer laptop.
|LM||Initial Release||Developer||License||Instruct / RLHF||Reference|
|BLOOM||2022-07-06||Hugging Face||Open RAIL-M v1||Link|
|GPTNeo||2021-03-21||EleutherAI, Together||Apache 2.0||Link|
|Llama 2||2023-07-18||Meta||Custom (Commercial OK)||Link|
|OpenLLaMA||2023-04-28||OpenLM Research||Apache 2.0||Link|
|Pythia||2023-02-13||EleutherAI, Together||Apache 2.0||Link|
|StableLM||2023-04-19||Stability AI||CC BY-SA 4.0||Link|
|Vicuna||2023-03-30||UC Berkeley, CMU, Stanford, MBZUAI, UCSD||Noncommercial||Link|
Commercial LLM Comparison
Side-by-side comparisons of different commercial LLM offerings.
Open Source LLM Comparison
Side-by-side comparisons of open source LLM options.
Scroll right to see the full table.
The most widely known LLMs are general-purpose, i.e. they can perform a variety of tasks across different topics and commercial industries. However, sometimes users and businesses may want an LLM trained on data from a specific industry, reducing the amount of prompting required for it to behave in an industry-relevant way and constraining its behavior. Also known as domain-specific LLMs, these language models may be easier to deploy to production for many businesses or serve as a better foundation for fine-tuning.
LLMs for biomedical, healthcare, finance, academia, and eCommerce.
LLMs are often trained on massive web crawls of text from various languages. Hence, often they are multilingual by default. However, there have also been LLMs trained specifically for languages besides English.
In addition to APIs, a number of developer libraries and SDKs have been released for working with LLMs. You can find Sapling's curated list of LLM libraries here:
Frequently Asked Questions
As these systems are evolving rapidly, we do not feel comfortable passing judgement on which LLM is best. However, a combination of cloud vs. ability to self-host, pricing, and qualitative evaluation should be enough to prune the index down to a small number of possible options.
If you'd like to look over tables of numbers, Stanford mantains the HELM benchmark.
Contact us with a brief description of your use case if you'd like for us to make a snap assessment. Depending on your requirements, a smaller, custom language model may even be the best option.
Please see the question above on how to evaluate different LLMs. Some factors you'll likely wish to consider include (1) compute costs, (2) data security requirements, (3) whether a custom language model would work best, (4) latency requirements, and (5) internal expertise available to set up the deployment.
LLMs are now available for different languages (Chinese, English, etc.) as well as different industries (healthcare/biomedical, legal, software coding, financial services, and cybersecurity). We plan to release comparisons for different languages and industries soon; in the meantime, feel free to contact us regarding your specific need.
Training an LLM is expensive. Although libraries and scaffolding for training LLMs are being rapidly released, the process can still be finicky, especially if you do not have experience training NLP models. If you need guidance on getting started, it's more than likely you should instead be finetuning one of the existing commercial LLMs using their finetuning guides and/or finding a LLM that roughly matches your use case.