Together AI has introduced Together Post-Training, a service for shaping open-weight language models into custom models a customer owns. The announcement is dated 10/19/2026. It supports training from supervised data (SFT) and from environment rewards (reinforcement learning).
What the service does
The customer defines the whole training loop: data, environments, agentic harnesses and extra steps such as logging or external benchmarking. Together says this is the difference from its existing fine-tuning service. Together runs the infrastructure underneath.
Experiments run on Model Resources, which are GPU allocations reserved for the customer. One resource can hold several Training Sessions, for example sets of LoRA adapters sharing the same GPUs. Full-weight training is available for select models. Each session has a Trainer for forward and backward steps and Generators for rollouts. For SFT, a resource can start with zero generators to save cost. Asynchronous RL is supported for long agentic rollouts.
Standard losses include cross-entropy, GRPO, CISPO and DPPO. Customers can also compute a custom objective locally and send the gradients to the Trainer.
Models and migration
Launch models are Muse Spark Open, GLM 5.3, Kimi K3 and Qwen 3.8-27B. Together says it will add more in the coming weeks. Teams that use the Tinker API can migrate by changing one import line, according to the company.
Deployment
Dedicated Model Inference now supports multi-LoRA adapter serving. Several trained adapters can be swapped within one GPU deployment rather than each needing its own. Together also says checkpoints keep their quality when served merged in FP4 precision.
Price
Billing is per GPU-minute. Together claims that at larger scale the total cost falls below token-based pricing. That is a vendor claim, not an independent measurement. For the first eight weeks, the price is $9 per GPU-hour of B200 usage.
Early users
The source quotes three early-access teams. Deel's Anthony Li says his team ran up to five concurrent RL experiments on the same GPU capacity. Jack & Jill says its model for predicting which candidates a hiring manager will want to meet matched a closed-source model in offline evaluations, using a cheaper open model. Collinear AI reports launching full-scale agentic RL runs through self-onboarding. These are customer statements published by the vendor.
Help from Together
Together's forward-deployed engineers can work with a customer across the lifecycle: choosing a model and harness, preparing data and training environments, running evaluations and shipping an optimized endpoint.