Together Post-Training: frontier-grade model shaping, the way that fits you
Together AI describes a way to turn open models such as Qwen, GLM and Kimi into custom models using SFT or RL: users write the training loop while Together manages the underlying infrastructure.
What the article says
Together AI published Together Post-Training on October 19, 2026, a service for building custom models through reinforcement learning with environment rewards and supervised fine-tuning on labeled data. Users define the training loop while Together handles infrastructure and optimization.
At launch the models are Muse Spark Open, GLM 5.3, Kimi K3 and Qwen 3.8-27B, with more promised in the following weeks. The service supports LoRA (including several adapters sharing GPUs) and full-weight training for select models, with cross-entropy, GRPO, CISPO and DPPO losses and asynchronous reinforcement learning. Billing is per GPU-minute, at $9 per B200 GPU-hour for the first eight weeks. Tinker API users can switch by changing one import line, and forward-deployed engineers help with model selection, data, evaluation and deployment. Deel and Jack & Jill are among the early customers cited.
Read the official article AnnuIA profile
Summary written by AnnuIA in its own words from the official page of Together AI. The content and rights belong to the publisher: please read the original article.