Together AI Launches Post-Training: RL and SFT on B200 at $9 per GPU-Hour

Together AI has opened Together Post-Training. The service lets users train their own models with RL, using rewards from an environment, or with SFT on labeled data, without building the infrastructure themselves. Four open models are available at launch: Muse Spark Open, GLM 5.3, Kimi K3 and Qwen 3.8-27B. For the first eight weeks, an hour of B200 time costs $9. The beta has been running since April, and the service is now open to everyone.

How the service works

Unlike Together's standard fine-tuning service, Post-Training lets the user define the entire training loop. You can plug in any data, environments and agent harnesses, and add custom steps such as logging or external benchmarks. Compute runs on Model Resources, which are dedicated GPUs that do not compete with other customers' requests. A single resource holds one or more Training Sessions, and each session corresponds to a separate experiment on one base model. Inside a session, a Trainer handles the forward and backward passes, and a Generator handles policy rollouts.

Key features of the service:

• modes: LoRA, including several adapters on a single GPU allocation, and full-weight fine-tuning for some of the models;
• built-in loss functions: cross-entropy, GRPO, CISPO, DPPO, plus the option to compute your own loss locally and send gradients to the Trainer;
• asynchronous RL to keep GPUs busy during long agentic rollouts;
• for SFT, a resource can run with zero generators, so you don't pay for capacity you don't use;
• billing by GPU-minute;
• migration from Tinker scripts by replacing a single import line: import together.post_training.tinker as tinker.

Where the speed and savings are hiding

The main technical risk in RL, according to the authors, is a mismatch between the training and inference engines. If the same weights produce slightly different outputs because of numerical differences, a run can stall or collapse. Together says it removes the sources of policy drift using standard methods such as Rollout Routing Replay, along with its own work on training stability.

Most of the time in RL goes to generating rollouts, and here Together relies on its inference stack. For multi-step agents, caching and routing cut redundant computation and keep memory in check. On the training side, optimized parallelism and sequence packing provide the speedup. The company claims that as experiments scale, the cost of training falls below token-based pricing. The article contains no independent measurements; this is the vendor's own claim.

The post-training stage has its own fix. Dedicated Model Inference now offers multi-LoRA serving: adapters can be swapped on a single GPU deployment instead of requiring a separate deployment for each run. Together also promises that checkpoint quality holds when moving from multi-LoRA sampling to a merged model in FP4.

What early customers say

Deel reports that real RL experiments were up and running within a couple of hours, with up to five parallel runs on a single GPU allocation. Jack & Jill fine-tuned a model that predicts which candidates a hiring manager will want to see. In offline evaluations it matched a closed model, but the open model is cheaper to serve. Collinear AI has used the service since the April beta and runs full-scale agentic RL jobs. All quotes come from Together itself.

Practical takeaways for teams

If you already have an agent running on a closed model and a clear quality metric, the service lets you test whether a specialized open model can replace it. The discount period makes entry cheap: $9 per B200 GPU-hour for eight weeks. To try out an idea, there is a cookbook on GitHub with examples of GRPO on GSM8K, DPO as a custom loss, and several LoRAs on a single resource.

For teams on Tinker, migration will take minutes. Before switching, compare your offline metrics across several base models and calculate the total cost of serving in FP4. For companies without RL expertise, Together offers forward-deployed engineers.

They help pick a model and agent harness, build data and environments, and then ship an optimized endpoint.