The author provides a brief overview of each model, including their launch dates, creators, unique features, and availability. For instance, Google's Gemma series, launched in February 2024, is optimized to run on consumer hardware and is available for testing on Kaggle and training and deployment through Google Cloud Vertex AI. On the other hand, SONAR, introduced by Meta in August 2023, is a multilingual and multimodal fixed-size sentence embedding space covering 200 languages. The author promises to provide more details on other growth vectors at the foundation model level in a subsequent post.
Key takeaways:
- The article discusses the recent growth in the development of new foundation models by AI companies and start-ups, driven by the need to build AI applications optimized for a variety of use cases, cost requirements, and usage ranges across Consumer and Enterprise verticals.
- The main categories of these models include Cost & Performance, Embeddings, Visual, Audio, 3D, General Purpose, Domain Focus, and Language Focus.
- Several models from notable AI companies have been launched in the past few months, including the Gemma series from Google, Stable LM 2 1.6B, Mixtral 8x7B, phi-2 from Microsoft Research, Orca2, Nemotron-3 8B from NVIDIA, Titan Text Lite from Amazon, Mistral 7B, MPT-7B-8k from MosaicML, Embedding-3 Small and Large from OpenAI, Embed v3 from Cohere, Titan Text Embeddings and Titan Multimodal Embeddings from Amazon, and SONAR from Meta.
- The article promises a Part 2 with more details on the other growth vectors at the foundation model level and invites readers to suggest more representative models for these categories.