Together AI launches managed post-training service for open models

Together AI said it launched Together Post-Training on October 19, expanding access to managed model customization beyond a select early-access group. The service supports supervised fine-tuning, which teaches a model from labeled examples, and reinforcement learning, which adjusts it using rewards from an environment. Both are available for selected open-weight models on dedicated computing resources.

The service debuted at a promotional price of $9 per B200 GPU-hour, down from a displayed standard rate of $11, through December 14, 2026. Because that is a per-GPU rate, a resource’s total hourly cost depends on its GPU count. Under the billing rules, charges accrue by the GPU-minute while a reserved Model Resource is ready, rather than by tokens processed. Queueing, startup and shutdown are not billed. An idle ready resource, however, continues accruing charges until it is stopped or automatically shuts down after one hour without an operation.

Customers define the Python training loop, data, rewards, environments and loss function; Together AI runs the trainer. For reinforcement-learning jobs, the company also handles rollout generation and weight synchronization. Its product page says the service supports synchronous and asynchronous workflows. Training takes place on dedicated “Model Resources” rather than shared request capacity, and one resource can host one or more sessions using the same base model.

The launch post named Muse Spark Open, GLM 5.3, Kimi K3 and Qwen 3.8-27B, while saying the selection would expand. The supported-model documentation, observed after launch, listed six public configurations: Qwen3.6 35B A3B, Qwen3.8 27B, Kimi K3, GLM 5.3, MiniMax M3 and Muse Spark Open 816B A42B. Together AI says all six support LoRA, a method that trains small added parameter sets instead of changing every model weight. The documentation lists full-weight training only for Qwen3.8 27B.

Together AI’s own materials disagree on one capability. The product page labels Qwen3.6 35B A3B as supporting full-weight training, while the supported-model documentation says it supports LoRA only. The documentation also says the service currently supports text models on B200 GPUs, with B300 and GB300 support planned for later.

The company says its work on numerical alignment, caching, routing, parallelism and sequence packing improves training quality, speed and utilization. It also says some scaled experiments can cost less than token-priced alternatives. The official materials supplied for this story do not include an independent benchmark for those claims. Together AI had previously introduced the underlying custom-training system in a limited closed beta on July 23.