Together

Together.ai hosts a cloud platform purpose-built for training and running open-source generative models.
Main introduction (from homepage hero and “Build gen-AI with open-source” block): the service gives teams instant access to 200+ OSS models—Llama-3, MixtraL, Stable Diffusion, DeepSeek, etc.—through a single GPU-optimized stack, so developers can move from prototype to production without building infrastructure.

Key features extracted strictly from the same page:

1. Serverless endpoints: auto-scaling, pay-per-token inference on the full model catalog; first 1 M tokens free each month.
2. Dedicated GPU clusters: on-demand A100/H100 nodes with one-click deployment and zero virtualization overhead, priced 2–3× below hyperscalers.
3. Fine-tuning & custom pre-training: built-in recipes for LoRA, full-weight updates, and continued pre-training on proprietary data; jobs launch in <30 s.
4. Inference optimization: custom CUDA kernels, FlashAttention-2, continuous batching, and token streaming that together deliver up to 4× higher throughput than standard Hugging Face pipelines.
5. Unified API & dashboard: one REST/WebSocket interface and Python SDK to switch between models, monitor usage, set rate limits, and audit costs in real time.