Together AI is a cloud platform for open-weight models, covering inference, fine-tuning, and dedicated GPU clusters. Access runs through a REST API with official clients; the Python library provides synchronous and asynchronous clients plus a command-line interface, and targets Python 3.10 or later.
The inference surface covers chat completions, text completions, image generation, embeddings, and reranking. Chat messages accept plain text as well as image and video URLs as content items, and requests support streaming, logprobs for token-level confidence, and a batch endpoint that returns larger jobs within 24 hours.
Fine-tuning runs through the same client: upload a JSONL training file, create a job with epochs, checkpoints, batch size, learning rate, and a suffix, list events while it runs, cancel it, and download the resulting weights or a checkpoint. Every one of these operations has a command-line equivalent.
Version 2 of the Python SDK is published separately as together-py, with a migration guide; version 1 is in maintenance mode and receives no new features.
Features
- Chat completions: requests name a model such as meta-llama/Llama-4-Scout-17B-16E-Instruct and return standard choice objects
- Multi-modal messages: text, one or more image URLs, and video URLs in a single message
- Streaming and async clients: token streaming plus an asynchronous client for concurrent requests
- Batch inference: file-based jobs with a 24-hour turnaround and status polling
- Embeddings and reranking: embedding models and a rerank endpoint that scores documents against a query
- Image generation: image models called with a prompt, step count, and sample count
- Fine-tuning: file upload, job creation, event listing, cancellation, and checkpoint download
- Logprobs: token-level generation probabilities returned alongside completions
- Model listing: an endpoint and command that enumerate the available models
- Command-line parity: chat, completions, image generation, files, fine-tuning, and models from the shell
