Together Post-Training opens flexible RL and SFT on open models

A typical hosted fine-tuning service takes a dataset and a few settings and hands back a trained model; the training procedure itself is fixed. Together AI's new Post-Training service hands that procedure to the customer. In its launch post, Together says the entire training loop is yours to define, in contrast to its existing fine-tuning product: you bring any data, reinforcement learning environment or agent harness, insert steps such as logging or outside benchmarks, and decide when each step runs, while Together handles the GPUs and the systems work underneath.

The service splits each experiment into two parts. A Trainer runs the forward and backward passes and updates the weights, and Generators produce rollouts, meaning the model's attempts at a task that a reward then scores in reinforcement learning. Both run on a Model Resource, a GPU allocation reserved for one customer rather than shared with other users' requests, which can host several experiments on the same base model at once, such as multiple LoRA adapters. LoRA trains small add-on weight matrices instead of the whole model, though Together also offers full-weight training for select models. For supervised fine-tuning a resource can run with no generators to save cost. Together supports standard losses from cross-entropy to GRPO, and lets customers compute a custom objective on their own machine and send only the gradients.

Together's pitch rests on correctness as much as convenience. It says a well-known way reinforcement learning runs go wrong is when the training and inference engines compute slightly different outputs from the same weights, which can degrade or collapse a run, and that it works to eliminate those sources of drift. It also says rollouts reuse the optimizations behind its inference products, since generating rollouts is where most RL time goes.

The obvious comparison is Thinking Machines' Tinker, an API that lets researchers write training loops locally and run them on remote GPUs (earlier Intel). Together names Tinker as a service users can migrate from, saying an existing script can switch over by replacing one line. Pricing is per GPU-minute rather than per token, with a launch rate of $9 per B200 GPU-hour for the first six weeks. Launch models include Muse Spark Open, GLM 5.3, Kimi K3 and Qwen 3.8-27B.