Together AI launches custom model training service

Together AI launches Together Post-Training, a service for customizing open weight models through reinforcement learning and supervised fine tuning, as detailed in a company blog post. Users can define their own training loops, data, environments, agent harnesses and evaluation steps while Together AI manages the infrastructure.

The service supports LoRA training and full weight experiments for selected models. Dedicated GPU allocations can run multiple training sessions, while separate trainer and generator instances support synchronous or asynchronous reinforcement learning. Users can choose standard loss functions such as GRPO, CISPO and DPPO, or send gradients from a custom objective.

Initial models include Muse Spark Open, GLM 5.3, Kimi K3 and Qwen 3.8-27B. Together AI is offering the service at a reduced $9 per B200 GPU hour for the first eight weeks.