Together AI launches custom model training with developer-controlled experiments

Together AI launched Together Post-Training on October 19, 2026, giving developers control over custom model-training experiments without operating the underlying GPU infrastructure. The service supports learning from labeled examples or rewards, while customers choose their training loops, data and environments—a different arrangement from Together’s existing, more prescribed fine-tuning service.

From early access to an open training loop

The public launch follows several months of work with early-access customers. In a testimonial published with the announcement, Collinear AI researcher Adit Jain said his team had worked with Together since the Post-Training beta began in April. Together says the service is now available for everyone to use.

Together frames the engineering burden as a reason companies stall on custom models. Training requires repeated changes to data, algorithms and evaluations, alongside infrastructure that becomes harder to coordinate as models grow. Its response is to manage that infrastructure while leaving the experimental choices with the customer.

Customers can bring their own data, environments and software for running agents, then add logging or external benchmarks anywhere in the training process. Supervised fine-tuning uses supervision data; reinforcement learning uses rewards from an environment. Both sit inside the same service rather than requiring separate infrastructure arrangements.

The launch lineup includes Muse Spark Open, GLM 5.3, Kimi K3 and Qwen 3.8-27B. Together supports LoRA adapter training, including multiple adapters sharing GPU capacity, and full-weight experiments for selected models. That last option is limited by model support, not available across the entire lineup.

Dedicated GPUs, separate training and generation

Experiments run on what Together calls Model Resources: dedicated GPU allocations that do not compete with other customers’ requests for priority. Each allocation can hold one or more training sessions using the same base model. Several LoRA experiments can therefore share a resource instead of reserving separate GPUs for every variation.

Within a session, a Trainer calculates training steps and updates model weights. Generator components produce the model responses used in reinforcement-learning experiments. Developers decide when updated parameters move from the trainer to the generator. For supervised fine-tuning, they can start a resource without generators to save money.

The service also supports asynchronous reinforcement learning, allowing generation and training to overlap. Customers can choose how many response-generating replicas an experiment needs. Beyond standard training objectives, they can calculate a custom objective locally and send its gradients—the signals used to adjust model weights—to Together’s trainer.

Coordination is not just a scheduling problem. Together warns that small numerical differences between training and response generation can weaken a run or cause it to collapse. It says its infrastructure reduces those mismatches using techniques including Rollout Routing Replay. That is Together’s engineering claim, not an independently measured result presented here.

$9 per GPU-hourB200 usage

Together offers the reduced B200 rate for the first eight weeks. Model Resources are billed per GPU-minute.

Taking an experiment into production

After training, Together’s Dedicated Model Inference service can serve multiple LoRA adapters within one GPU deployment. Teams can swap adapters to compare results without creating a separate production deployment for each experiment. Together also says its optimizations preserve checkpoint quality when moving to faster serving configurations, including lower-precision FP4 execution.

One early customer describes a narrower, task-specific payoff. Jack & Jill founding engineer Oliver Normand said the company trained a model to predict which job candidates a hiring manager would want to meet. In offline evaluations, he said, it matched a closed-source model’s quality while being significantly cheaper to serve. The testimonial supplies neither a named comparison model nor numerical results.

Developers can begin with the latest Together SDK and the company’s post-training cookbooks. Examples cover reward-based training on the GSM8K benchmark, implementing a custom training objective and running multiple adapters on one resource. Together also offers a Tinker migration path that it describes as replacing one import line while leaving the rest of a script unchanged.

The next promised change is broader model coverage: Together says it will expand the selection in the coming weeks. For companies seeking more than infrastructure, its forward-deployed engineers also offer help with model selection, data, environments, evaluation and deployment. That assistance is available alongside the self-directed service, rather than replacing customers’ control of their experiments.