Not-so-large language models: a concise guide

Meta, Mistral, Cohere, and Databricks all have impressive open-weight models.

When OpenAI released GPT-4 in March 2023, it set a new standard for large language model performance. Since then, two companies—Google and Anthropic—have created models in the same league as GPT-4.

Many other companies have focused on building smaller models. Small models are useful because a model’s parameter count drives inference costs. GPT-4 is more …