Together launches Post-Training service for custom AI models

Together introduced Together Post-Training on Oct. 19, 2026, opening its custom-model training service beyond the organizations it says had early access over the preceding months. The company says developers can use supervised fine-tuning with supervision data or reinforcement learning with environment rewards while defining their own training loops. Together handles the underlying infrastructure, and a separate launch post says the service is live.

Developers can choose their data and training environments and add steps such as logging or external benchmarking, according to Together. Its Model Resources are dedicated GPU allocations that can hold multiple Training Sessions using the same base model. Together says multiple LoRA adapter experiments can share GPU capacity, while full-weight experiments are available for select models. Its launch thread also points users to cookbook recipes.

Together lists Muse Spark Open, GLM 5.3, Kimi K3 and Qwen 3.8-27B among the models available at launch and says it plans to expand the selection. After training, the company says its Dedicated Model Inference lets developers swap LoRA adapters in a single GPU deployment to compare outputs.