Together Post-Training: frontier-grade model shaping, the way that fits you

Together AI has launched Together Post-Training, a new service for building custom models on its platform with either reinforcement learning from environment rewards or supervised fine-tuning from labeled data. The pitch is simple: let teams shape models without having to build and babysit the whole training stack themselves.

That matters because custom model training is no longer a side project. Together says companies are using it to cut costs, improve task-specific performance, and move beyond prompt tricks and harness tweaks. The company is aiming at the awkward middle ground where teams want frontier-level quality, but not frontier-lab headcount.

The service gives users control over the training loop while Together handles the machinery underneath. It supports LoRA training, including multiple adapters sharing one GPU allocation, and also full-weight experiments for select models. Users can plug in their own data, environments, agentic harnesses, logging, and external benchmarks. Model Resources are dedicated GPU allocations, and a single resource can hold one or more Training Sessions.

Together says the system is built to keep output quality, speed, and cost in line. It points to protections against drift between training and inference, rollout optimizations drawn from its serving stack, and support for async RL when long agent rollouts would otherwise waste GPU time. For SFT, a model resource can run with zero generators to save money. Pricing is per GPU-minute, and for the first eight weeks the company is charging $9 per GPU-hour for B200 usage.

The launch starts with open-weight models including Muse Spark Open, GLM 5.3, Kimi K3, and Qwen 3.8-27B, with more coming. Together is also pushing a migration path for teams already using post-training tools like Tinker API, and it says its forward-deployed engineers will work alongside customers on model selection, data, evaluation, and deployment.