Overview
Together AI is a cloud platform built specifically for open-source AI. Instead of locking you into a single proprietary model, it gives developers a single OpenAI-compatible endpoint to run, fine-tune, and serve more than 200 open models - Llama, DeepSeek, Qwen, Mixtral, FLUX, and more. The product stack has three layers: Serverless Inference for pay-per-token prototyping with no infrastructure to manage; Dedicated Endpoints for reserved GPU capacity with SLA-backed latency; and GPU Cloud for raw H100/H200/B200 access when you want to train or self-host. Fine-tuning is first-class: upload a dataset, pick a base model, and launch a LoRA or full fine-tune through the API or dashboard, then deploy it straight back to serverless. A custom inference engine delivers some of the best throughput on NVIDIA GPUs - Llama-class models clear 200 tokens/second - which is why agent frameworks and multi-agent systems that make thousands of calls per task gravitate to it. New accounts get free credits, and batch mode cuts serverless prices 50% for asynchronous jobs. The trade-off is that, unlike Hugging Face or Replicate, the model catalog is curated rather than exhaustive, and the platform is text/code-first with thinner coverage of video and audio.
Key Features
- 200+ open models - One endpoint for Llama 4, DeepSeek-R1, Qwen, Mixtral, FLUX, and more, all OpenAI-compatible.
- Serverless inference - Pay only for tokens; no servers to manage; batch mode at 50% off for async workloads.
- Fine-tuning pipeline - LoRA and full fine-tunes via API or dashboard, deployable straight to serverless or dedicated endpoints.
- Dedicated endpoints - Reserved GPUs from $0.85/hr with no cold starts and guaranteed latency for production.
- GPU Cloud - Raw H100/H200/B200 access for custom training and self-hosting.
- Transparent pricing - Per-token rates known up front; no GPU-second math or hidden credit pools.
Pricing
| Plan | Price | For |
|---|---|---|
| Free credits | $1-$5 signup | Try models before paying; 71+ models free |
| Serverless | From $0.06/M tokens | Pay-per-token prototyping and production (Llama 3.1 8B $0.06-$0.18/M) |
| Dedicated endpoints | From $0.85/hr | Production workloads needing reserved capacity and SLA |
| GPU Cloud | From $5.49/hr (H100) | Custom training and self-hosting on NVIDIA GPUs |
Comparison
vs. OpenRouter: OpenRouter is a unified gateway to many providers’ models with a similar drop-in API, but Together AI owns its inference stack and pairs it with integrated fine-tuning and dedicated GPU capacity. Choose OpenRouter for breadth across vendors; choose Together AI when you need to customize and serve your own open models.
vs. Replicate: Replicate shines for media generation - image, video, audio, 3D - through a huge community model catalog. Together AI is text/code-first and wins on LLM throughput and fine-tuning, so it is the better default for agent and RAG backends.
vs. Hugging Face: Hugging Face is the model hub and community; Together AI is the inference and training runtime. Many teams browse models on HF, then run and fine-tune them on Together AI for production speed.