Together AI is a cloud platform for running, training and serving open-source AI models. It hosts models for chat and text, code, images, video and voice, and rents GPU clusters (groups of graphics processors, the chips that run AI models) for training and large batch jobs. The company also publishes systems research, such as FlashAttention.
A program sends requests to Together's API, which is compatible with the OpenAI API, so code written for OpenAI can switch by changing the address and the key. Serverless inference runs a model on demand with no servers to manage; batch inference processes large jobs asynchronously; dedicated inference reserves hardware for one model. For speech-to-text, Together hosts OpenAI's Whisper and NVIDIA Parakeet models: a program sends an audio file or a URL and receives the transcript, either as a batch job or as a real-time stream. Fine-tuning trains a model on a customer's own data and then deploys it.
Billing is per use: per million tokens for language models, per audio minute for transcription, per image or video, and per GPU hour for dedicated hardware and clusters.