Together AI Launches Post-Training Service With Reserved GPU Capacity

Together AI launched Together Post-Training on Oct. 19, giving developers reserved GPU capacity to customize open-weight AI models through reinforcement learning and supervised fine-tuning while it manages the underlying infrastructure.

Developers choose their training data, objectives and experiment steps. The service supports multiple LoRA training experiments on the same reserved GPU capacity, as well as full-weight training for selected models and asynchronous reinforcement learning. Post-training changes a model after its initial development.

The initial lineup includes Muse Spark Open, GLM 5.3, Kimi K3 and Qwen 3.8-27B. Together AI is offering an introductory rate of $9 per B200 GPU-hour for the first eight weeks.

Customers can also serve several trained LoRA adapters through one dedicated inference deployment and compare the resulting model variants. Inference is the stage when a deployed model processes a request and generates a response. The service supports serving trained checkpoints in FP4 precision.

Deel was among the early users. Anthony Li, its associate director of data science, said the service allowed “up to five concurrent RL experiments on the same GPU capacity.”

The launch follows Together AI’s $800 million Series C financing on July 1 and commitments for more than 500 megawatts of independently financed computing capacity. Together AI plans to add more models in the coming weeks.