Cerebras Systems, an artificial intelligence chip startup, has released seven open-source models trained on its AI supercomputer, Andromeda. The move is intended to promote greater collaboration and to provide research and business communities with tools for developing AI applications.
The seven models range from 111 million parameter language models to 13 billion parameter models. These models can be run on any hardware and give users greater access to AI tools.
According to Cerebras CEO Andrew Feldman, the release of these models is part of a larger movement to keep AI open source, given the industry’s huge money. Smaller models can be run on phones or smart speakers, while larger models require PCs or servers to run, though larger models are required for more complex generative functions.
Cerebras’ larger model trained in just over a week, a process that normally takes months. This was made possible by the Cerebras system’s architecture, which includes a chip the size of a dinner plate designed for AI training. Cerebras GPT is the first GPT model family that is compute-efficient at all model sizes. Existing open GPT models are trained on a predefined set of data tokens. Cerebras GPT establishes a new high-accuracy baseline for broad use by applying the Chinchilla training recipe across all model sizes.
The release of these models is expected to increase market competition and help more startups like Cerebras gain market share. While larger models aren’t always better, Cerebras claims that its models are compute-efficient at all sizes, resulting in more accurate and effective models for users. It will also provide new opportunities for businesses and researchers to develop AI applications.
The sources for this piece include an article in Reuters.
