Together AI launches Post-Training service for custom models

News

Together AI introduced Post-Training for RL and SFT workflows, with dedicated GPU resources, multi-LoRA support and introductory B200 pricing.

Together AI on 2026-10-19 introduced Together Post-Training, a service aimed at helping developers build custom models using reinforcement learning and supervised fine-tuning while the company manages the underlying infrastructure, according to the company’s blog post published that day.

Evidence and context

The launch is positioned as a way to let users define their own training loops while Together handles the GPU systems underneath, including Model Resources, Training Sessions, Trainer instances and Generator instances, the company said. Together said the service supports LoRA training, multiple LoRA adapters sharing the same GPU capacity, and full-weight experiments for selected models, while also allowing asynchronous RL and custom training objectives.

Together said the service launches with open-weight models including Muse Spark Open, GLM 5.3, Kimi K3 and Qwen 3.8-27B, with more models to be added in the coming weeks. The company also said it is offering the service at a reduced price of $9 per GPU-hour of B200 usage for the first eight weeks.

The announcement comes after Together said on 2026-07-01 that it had raised an $800 million Series C and secured commitments for more than 500 MW of independently capitalized compute capacity, providing funding context for its infrastructure expansion. It also follows a 2026-09-29 announcement with Mubadala about exploring UAE AI infrastructure opportunities and an Abu Dhabi presence.

Together said the service is designed to reduce the infrastructure burden involved in post-training, improve rollout and training performance, and make trained checkpoints easier to deploy or refine. The company’s blog also said the offering includes support for multi-LoRA adapter serving in Dedicated Model Inference and optimization for serving checkpoints in FP4 precision. Those capabilities are company-reported and were not independently benchmarked in the provided material.

In its launch note, Together included testimonials from early-access customers at Deel, Jack & Jill and Collinear AI describing faster experiment setup and the ability to run RL workloads on shared GPU capacity. Those comments were attributed by the company and are not independently verified in the source material.