Together AI launches post-training service with custom training loops

Together AI launched Together Post-Training, where users build custom models via SFT supervised data or RL environment rewards. The service offers two levels of abstraction: Model Resource (dedicated GPU allocation) and Training Session. Each session includes a Trainer instance handling forward/backward passes and weight updates, plus a Generator instance running policy rollouts. It supports LoRA, multiple LoRA adapters sharing GPUs, and full-weight training for some models; built-in losses include cross-entropy, GRPO, CISPO, and DPPO, and users can compute custom losses locally and send gradients back to the Trainer. First-wave models: Muse Spark Open, GLM 5.3, Kimi K3, Qwen 3.8-27B. Pricing is per GPU-minute, with B200 at a promotional $9 per GPU-hour for the first eight weeks. For SFT, setting Generator count to 0 cuts costs, and multi-LoRA experiments sharing one Model Resource let parameter sweeps avoid reserving separate GPUs per configuration. For agentic RL teams, async RL plus caching and routing optimizations for multi-turn rollouts reduce waiting; after training, multiple LoRA adapters can be swapped within a single-GPU deployment to compare results, and checkpoints can be merged for FP4 deployment. Users with existing Tinker API scripts can migrate by changing one import line.