Comparing models in different use cases

There is a long list of models available in Stack AI, but before getting too overwhelmed with the description of each model, let’s discuss some of the key use cases and which models are preferred.

USE CASEDESCRIPTIONPREFERRED MODELS
Formatting a promptSometimes an LLM can improve the prompt given by the user before sending it to another LLM that will perform the task. In this case, a lighter, less costly and faster model is preferred.gpt-3.5-turbo, davinci, claude-instant-v1
Summarizing a large document set or websitesIn this case, receiving broader context is crucial for the LLM to understand the overall meaning of the document and to summarize it correctly. A model with larger context window is preferred.claude-3-opus
Performing complex tasksImagine an LLM receiving input from the user, data from a list of documents, instructions as to how to behave, and needs to use an external tool to retrieve additional information to finally answer the user. In this case, the more powerful the best.gpt-4o
Performing complex tasks requiring large contextSame as the use case above, but requiring a larger context window to read entire documents. To deploy this use case at scale, it is preferred to use a well-trained, large window type of model that has been around for some time.claude-3-opus, gpt-4o

List of all available LLMs

GPT 4

GPT 4 is a large multimodal model (accepting text inputs and emitting text outputs today, with image inputs coming in the future) that can solve difficult problems with greater accuracy than any of our previous models, thanks to its broader general knowledge and advanced reasoning capabilities.

MODELDESCRIPTIONMAX TOKENSTRAINING DATA
gpt-4More capable than any GPT-3.5 model, able to do more complex tasks, and optimized for chat.8,192 tokensUp to Sep 2021
gpt-4-32kSame capabilities as the base gpt-4 mode but with 4x the context length.32,768 tokensUp to Sep 2021
gpt-4-turbo-previewBetter capabilities than gpt-4 and gpt-4-32k. Large context window and quick inferences.128,000 tokensUp to April 2023
gpt-4oThe latest and most advanced GPT model from OpenAI. Ideal for complex tasks requiring long contexts.128,000 tokensUp to Oct 2023

GPT-3.5

GPT-3.5 is a mid-generation upgrade of GPT-3 with fewer parameters. It includes a fine-tuning process that involves reinforcement learning with human feedback, which helps to improve the accuracy of the responses.

MODELDESCRIPTIONMAX TOKENSTRAINING DATA
gpt-3.5-turboMost capable GPT-3.5 model and optimized for chat at 1/10th the cost of text-davinci-003. Will be updated with the latest model iteration 2 weeks after it is released.4,096 tokensUp to Sep 2021
gpt-3.5-turbo-16kSame capabilities as the standard gpt-3.5-turbo model but with 4 times the context.16,384 tokensUp to Sep 2021

Anthropic

Anthropic’s model called Claude is a is a transformer-based LLM, much like GPT-3, that leverages large-scale machine learning techniques. The model is trained on a diverse range of internet text, giving it the ability to generate text that is coherent, contextually relevant, and remarkably human-like.

Below the comparison between the different versions of Claude.

LATEST MODELDESCRIPTIONMAX TOKENS
claude-3-opusClaude 3 has the best perofrmance performance, longer responses. Largest context window available in the market.200,000 tokens
claude-3-sonnetFaster than OpenAI GPT 4 and almost as good. Largest context window available in the market.200,000 tokens
claude-3-haikuLighter, less expensive, and much faster option.200,000 tokens

Google

Stack AI has early access to Google’s PaLM 2 model, the Large Language Model (LLM) released by Google. It is highly capable in advanced reasoning, coding, and mathematics. It’s also multilingual and supports more than 100 languages. PaLM 2 is a successor to the earlier Pathways Language Model (PaLM) launched in 2022.

The two models available are the following.

LATEST MODELDESCRIPTIONMAX TOKENSTRAINING DATA
gemini-1.5-proFine-tuned to follow natural language instructions and is suitable for a variety of language tasks1 million tokensUp to Feb 2024
gemini-proFine-tuned to follow natural language instructions and is suitable for a variety of language tasks30,720 tokensUp to Feb 2023
text-bison-001Fine-tuned to follow natural language instructions and is suitable for a variety of language tasks8,192 tokensUp to Feb 2023
chat-bison-001Fine-tuned for multi-turn conversation use cases.4,096 tokensUp to Feb 2023

Meta

LATEST MODELDESCRIPTIONMAX TOKENSTRAINING DATA
Llama-3-70b-chatLlama-3 is a state-of-the-art large language llm designed for enhanced reasoning, coding, and broad application across multiple languages and tasks.8,000 tokensUp to March 2023
LLama-3-8b-chatA smaller version of llama-3 that allows for faster inference.8,000 tokensUp to March 2023
Llama-2-70b-chatFine-tuned to follow natural language instructions and is suitable for a variety of language tasks4,096 tokensUp to Sep 2022
Llama-2-13b-chatA smaller version of llama-2 that allows for faster inference.4,096 tokensUp to Sep 2022

Mistral

LATEST MODELDESCRIPTIONMAX TOKENSTRAINING DATA
mistral-largeOur flagship model that’s ideal for complex tasks that require large reasoning capabilities or are highly specialized (Synthetic Text Generation, Code Generation, RAG, or Agents).32,000 tokensUp to Dec 2021
mistral-mediumIdeal for intermediate tasks that require moderate reasoning (Data extraction, Summarizing a Document, Writing emails, Writing a Job Description, or Writing Product Descriptions)32,000 tokensUp to Dec 2021
mistral-smallSuitable for simple tasks that one can do in bulk (Classification, Customer Support, or Text Generation)32,000 tokensUp to Dec 2021
mistral-8x22b-instructA 22B sparse Mixture-of-Experts (SMoE). Uses only 39B active parameters out of 141B.64,000 tokensUp to Dec 2021
mistral-8x7b-instructA 7B sparse Mixture-of-Experts (SMoE). Uses 12.9B active parameters out of 45B total.32,000 tokensUp to Dec 2021
mistral-7b-instructMistral’s very first model. A 7B transformer model. Small, yet very powerful for a variety of use cases.32,000 tokensUp to Dec 2021

Perplexity

The perplexity node lets you use Perplexity’s RAG fine tuned models. These models are ideal when you need to do in real time web search within your workflow.

MODELDESCRIPTIONMAX TOKENSTRAINING DATA
llama-3-sonar-large-32k-onlinePerplexity’s most powerful LLM model. Built on top of llama-2-70b.32,000 tokensOnline model
llama-3-sonar-small-32k-onlineA smaller but faster LLM that is built on top of mistral-7b.32,000 tokensOnline model

TogetherAI

TogetherAI provides a high-performance inference engine that ensures rapid processing speeds for large language models (LLMs). Renowned for its exceptional performance, scalability, seamless integration, and extensive support services, TogetherAI stands out in the industry. With StackAI, you can leverage the robust TogetherAI infrastructure through the TogetherAI node, accessing a diverse range of models from various families, including Mistral, LLaMA, Snowflake, and more.

Groq

Groq leverages custom hardware and infrastructure to deliver high-speed inference for a variety of large language models (LLMs). This allows for efficient processing, making it an ideal choice for applications requiring low latency and high throughput. Among the models available in this node, you will find llama3-70b, llama3-8b, mixtral-8x7b, gemma-7b.

Azure

Microsoft Azure offers the ability to host private clouds with OpenAI models. You can add these models in Stack AI using an “Azure” node. Hosting models in Azure has a few benefits:

  1. Lower and consistent latency: models in Azure are not affected by the traffic of OpenAI and

  2. Higher rate limits: models in Azure offer higher rate limits of up to 240,000 tokens per minute and 1440 requests per minute.

  3. Data privacy and compliance: data sent to Azure is kept under the private cloud and is not sent to OpenAI or any external service. These models are covered under Azure’s Business Associate Agreement (BAA) and are HIPPA compliant.

Enterprise users of Stack AI have access to models hosted in Azure.

Bedrock

Access a wide variety of LLMs from different providers hosted on AWS Bedrock. You may also provide your own API keys and use your own models hosted on your VPC.