The Pulse
Together AI Launches Custom Post-Training for Open Models
Together AI introduced a post-training service that lets developers define training workflows while the company provides GPU resources and infrastructure. Its launch announcement also describes supported models, a temporary B200 price and r
AI.info Team · Written with AI · fact-checked ·
AI teams face a trade-off in post-training: they can keep control of the training process, or avoid building and maintaining the infrastructure that makes it run. On October 19, Together AI introduced a service designed to offer both: customers define the training loop and choose their data and rewards, while Together supplies the GPU resources and training systems. The company says the service is available to all customers following several months of early access.
Together gives developers the training loop
Together Post-Training supports reinforcement learning from environment rewards and supervised fine-tuning from labeled data. Developers can use their own environments and agent frameworks, add logging or outside benchmarks, and choose from standard training objectives including cross-entropy, GRPO, CISPO and DPPO. They can also calculate a custom objective locally and send its gradients to Together’s Trainer.
The service offers LoRA experiments, including multiple adapters sharing GPU capacity, and full-weight training for selected models. Together says customers’ GPU allocations are dedicated to their experiments rather than shared with other users’ requests. Each experiment pairs a Trainer, which updates model weights, with Generator instances that produce rollouts; teams can choose synchronous or asynchronous reinforcement learning and omit generators for supervised fine-tuning.
Together’s launch lineup includes Muse Spark Open, GLM 5.3, Kimi K3 and Qwen 3.8-27B. The company says it plans to add models in the coming weeks. It bills resources by GPU-minute and is offering B200 usage at $9 per GPU-hour for the first eight weeks after launch.
The pitch: custom models without a custom training stack
Managing the infrastructure is the other half of the product. Together says its systems align training and inference computations to limit policy drift, and apply optimizations to rollout generation, parallelism and sequence packing. For multi-turn agents, the company says caching and routing reduce repeated computation. Those are vendor descriptions of the service, not published comparative benchmarks in the announcement.
After training, customers can compare LoRA adapters through a single dedicated GPU deployment rather than starting a separate deployment for each experiment, according to Together. The company also says checkpoints can move into production inference at lower precision, including merged FP4 inference, while preserving quality. That matters to the product’s central promise: a training run should yield a usable model, not just a checkpoint that needs a separate deployment project.
Early customers report faster starts
Together’s launch post includes testimonials from early users, including Deel.
“The service is efficient, allowing us to run up to five concurrent RL experiments on the same GPU capacity.”
— Anthony Li, Associate Director of Data Science at Deel. Li also said Together’s research team gave Deel feedback on model selection, training techniques, reward functions and training efficiency.
Jack & Jill, a recruiting company, used the service to train a model that predicts which candidates a hiring manager might want to meet. Founding Engineer Oliver Normand said the company matched a closed-source model’s quality in offline evaluations using an open-source model that was cheaper to serve; the post gives no benchmark details or cost comparison. Collinear AI researcher Adit Jain said the company had worked with Together’s post-training beta since April and used the latest version to start agentic reinforcement-learning runs.
Control still comes with a provider dependency
The service does not remove the tension that opened the announcement: customers control their training recipes, but Together runs the infrastructure and defines the available model and hardware options. Full-weight training is limited to select models, and the launch price applies specifically to B200 GPU usage during the introductory period. Teams weighing the service will need to test how its supported models, resource billing and deployment workflow fit their own experiments.
Together also provides cookbooks for GRPO training on GSM8K, custom DPO losses and parallel LoRA experiments. Its migration guide offers a one-line import change for users of the Tinker API. The announcement’s clearest concrete customer result is Jack & Jill’s reported offline evaluation; Together does not publish the evaluation setup or numerical score.