Together AI

Together AI is an open-weights platform that still behaves like a research lab: a catalog of 200+ checkpoints across chat, code, image, video, audio, embeddings and rerank behind a single API, served by kernels the company writes itself. Deployment is a ladder you climb without switching vendors — serverless calls priced per token, batch endpoints at reduced rates, provisioned throughput units when capacity must be guaranteed, dedicated model and container endpoints for your own weights or LoRA adapters, and Together GPU Cloud for bare metal, self-serve from a few nodes to clusters of thousands with on-demand B200s and no egress charge on storage. Fine-tuning is first-class: LoRA or full-parameter over supervised or preference data, priced by dataset size with a base job fee from $4.

What it is not is the home of the frontier — the best model on the menu is an open one, so if your bar is whatever the closed labs shipped this quarter, a first-party API or an aggregator gets you closer. The win is legible economics: serverless rates from about a dollar per million tokens on mid-size models up to roughly $15 for the heaviest generation, dedicated GPUs from around $4 to $9 per hour, which turns rent-versus-self-host into arithmetic. Before committing, check what a model page actually prices (quantization and context window), the minimum units and term on provisioned throughput, and whether a fine-tune autoscales on shared capacity or bills hourly on dedicated hardware.