Introduction
This isn’t meant to be a comprehensive list of Large Language Models and hosts, rather it lists models I’ve used and where online they were hosted. When I started writing scripts that ran prompts against multiple models, I found it difficult to find the models I wanted to use - or at least to find where they were hosted online.
Models
Open AI GPT3.5, GPT4 & GPT4 Turbo
Pretty obvious entry for the list. You can apply for API access, but it’s not guaranteed. It does seem to be more available than it was when GPT3.5 and GPT4 were new.
Open AI has different specialized versions of its models for different applications, such as versions for working with Tools/Functions, more up-to-date versions trained on more recent data or versions with larger content windows. Unfortunately, not all of these are generally available through the API interface.
Open AIs full list of models is pretty extensive - here is the highlights as of November 2023:
| MODEL | DESCRIPTION |
|---|---|
| GPT-4 and GPT-4 Turbo | A set of models that improve on GPT-3.5 and can understand as well as generate natural language or code |
| GPT-3.5 | A set of models that improve on GPT-3 and can understand as well as generate natural language or code |
| DALL·E | A model that can generate and edit images given a natural language prompt |
| TTS | A set of models that can convert text into natural sounding spoken audio |
| Whisper | A model that can convert audio into text |
| Embeddings | A set of models that can convert text into a numerical form |
| Moderation | A fine-tuned model that can detect whether text may be sensitive or unsafe |
| GPT base | A set of models without instruction following that can understand as well as generate natural language or code |
Links
Anthropic Claude-Instant and Claude-2
Anthropic has Claude-Instant and Claude-2, positioned as direct competitors to GPT3 and GPT4. The performance of the models is very good, though they specialize in applications like chat interfaces and summarization. I haven’t seen anything from them to rival Open AI’s specialized model versions for Functions.
Anthropic models also have a huge context window of 100k tokens, which means you can do summarization without needing to use complex strategies to process the whole of large documents. On its own that makes them worth experimenting with.
Unfortunately, API access is still quite restricted - you can apply but there is no guarantee of being accepted.
Links
Meta AI Llama models
Meta AI has released its model as Open Source. This is a tremendous resource for the community, allowing experiments with fine-tuning a powerful model made with billions of parameters. Unfortunately, it also means that to work with it you either need to host it yourself or find someone who has made it available for public use.
I found DeepInfra perfect for my purposes, as they have the 7 billion, 13 billion and 70 billion parameter Llama models available. They also host a wide variety of other open-source models, including text generation, image generation and speech recognition models.
They also have a chat interface available with a selection of open-source text generation models available for use.
Links
Cohere Command
Command is Cohere’s flagship model, and is available through their API. It’s also available through their Coral chat interface.
Command is a powerful model, not quite in the league of Open AI or Anthropic models, but able to compete or even outdo Meta’s Llama models. I include it in model comparisons because it’s a good benchmark to show how well the more sophisticated models perform.
Links
Others
Trying to stay on top of developments in LLMs is almost impossible due to the incredible pace the field is moving. I haven’t listed HuggingFace on here, despite having used models hosted there before. They have a huge selection of models available, but only host some of the lower-powered versions and the performance of their API is poor. Having got a lot of errors from it I dropped it and swapped to using DeepInfra instead.
HuggingFace does have a lot of data sets and other resources though, so it’s worth checking out.
Conclusion
There are so many models available now, many of which I’d love to try out. I’m keeping the scope of this list to models I’ve used though, as I know all of these work reliably.