Together AI

Together AI is the general-purpose home for open-weight models. More than 200 of them — Llama, DeepSeek, Qwen, Kimi K3 and the rest — sit behind one API, with no subscription and no minimum spend. Together also rents the raw GPUs if you would rather run the whole thing yourself.

What is Together AI?

Together AI is a cloud platform for developers who want open-weight models without managing infrastructure. It sells four distinct things. Serverless inference is pay-per-token, from roughly $0.10 to $9 per million tokens depending on model size, with input priced below output. Dedicated inference puts a model on private GPU instances with guaranteed performance, for teams optimising latency at scale. GPU clusters rent raw NVIDIA hardware — 8 to 4,000+ GPUs, InfiniBand-connected — on demand or reserved up to six months, for teams training their own models. Fine-tuning is billed per token of training data, scaled by model size and method (LoRA, full fine-tune or DPO), with the resulting model landing on its own dedicated endpoint and per-hour meter.

Key Features

• 200+ models, one API: Swap between open-weight families without changing integration code.
• Batch API: Up to 30 billion tokens per model processed asynchronously at up to 50% below synchronous pricing.
• Fine-tuning: LoRA, full fine-tune and DPO on your own data, priced per training token.
• GPU clusters: From 8 to 4,000+ GPUs for teams running their own training or inference engines.
• Consumption billing: No subscription tiers, setup fees or minimum commitments.

Who Should Use Together AI?

Together suits teams building on open weights who want optionality — the ability to try Qwen against DeepSeek against Kimi K3 without three integrations — and teams that need to fine-tune on proprietary data. The Batch API makes it particularly attractive for offline workloads like classification, enrichment and evaluation, where halving the bill matters more than latency. If you need frontier closed models, you will still need an OpenAI or Anthropic account alongside it. If you only need one model as fast as possible, Groq is cheaper and faster on that narrow catalogue.

Pricing

Billing is pure consumption: you pay per token, per image or per video, with no subscription tiers, setup fees or minimum commitments. Serverless inference ranges from about $0.10 to $9 per million tokens depending on model size, with asymmetric input and output rates. Dedicated inference is billed per hour of GPU time. GPU clusters are on-demand or reserved up to six months. Fine-tuning is billed per token of training data and scaled by model size and method, with the resulting model on its own per-hour endpoint. The Batch API runs up to 50% cheaper than synchronous serverless.

Pros and Cons

Bottom Line

Together AI is the most flexible way to build on open weights: broad model choice, real fine-tuning, and a path from serverless calls all the way to renting your own cluster. Use the Batch API aggressively for anything that does not need to be real-time — halving the bill is the single biggest lever on the platform. Pair it with a frontier API for the reasoning-hardest parts of your product.