Together AI launches Post-Training service for custom RL and SFT models

October 19, 2026Source: Together AI

Together AI has introduced Together Post-Training, a service for building custom models through reinforcement learning (RL) with environment rewards and supervised fine-tuning (SFT). According to the company, developers define the training loop while Together handles the systems and infrastructure underneath.

Key points

Together Post-Training supports RL with environment rewards and SFT, with the full training loop defined by the user.
Experiments run on dedicated Model Resources, with LoRA (including multiple adapters per GPU allocation) and full-weight training for select models.
Launch models include Muse Spark Open, GLM 5.3, Kimi K3 and Qwen 3.8-27B, with more promised in the coming weeks.
Billing is per GPU-minute, and Together claims costs fall below token-based pricing at scale.
Multi-LoRA serving is now available in Dedicated Model Inference, with FP4 merged inference supported.

A training loop you define yourself

Together says the service is meant to give flexible and efficient model shaping without the heavy infrastructure burden. In contrast to its existing fine-tuning service, the entire training loop is left to the user. Developers can use any data, environments or agentic harnesses, and add steps such as logging or external benchmarking along the way.

The company argues that custom models are no longer a niche topic in 2026. According to Together, many AI-native and AI-forward organizations are training specialist agents that match the quality of frontier models at a fraction of the cost and runtime. Owning the resulting checkpoint also brings flexibility: it can be deployed in any location, further refined, or used locally.

Together points to the difficulty of post-training as the reason many companies never start or stall. It cites repeated iteration on data curation, training algorithms and evaluation protocols, plus increasingly complex RL infrastructure. Its forward-deployed team partners with customers on the end-to-end workflow.

How it works: Model Resources, Trainers and Generators

All experiments run on Model Resources, which are GPU allocations that Together says are fully the customer's and do not compete with other users' requests. Each Model Resource can hold one or several Training Sessions, and each session corresponds to one experiment on the same base model, for example one set of LoRA adapters.

Users can run LoRA training, including multiple LoRA adapters sharing the same GPU capacity, as well as full-weight experiments for select models. Each session has a Trainer instance that runs forward and backward steps and updates weights, and Generator instances that execute rollouts from the policy. For SFT, a model resource can be started with 0 generators to save costs. Users decide when trainer parameters are sent to the generator, and the service also supports asynchronous RL for high GPU utilization, especially with long agentic rollouts.

Supported standard losses range from cross-entropy and GRPO to CISPO and DPPO. Users can also compute a custom objective locally and send its gradients to the Trainer.

The service launches with several open-weight models, including Muse Spark Open, GLM 5.3, Kimi K3 and Qwen 3.8-27B. Together says it will expand the selection in the coming weeks.

Quality, speed and cost

Together says it made output model quality its north star. It highlights mismatched numerics between training and inference engines as a notorious RL problem, since the same weights producing slightly different outputs can make a run underperform or collapse. The company says it removes sources of policy drift using standard techniques such as Rollout Routing Replay and further optimizations from its own research on training stability.

On speed, Together notes that most RL time goes to generating rollouts, so its rollout engines reuse optimizations from its serving stack. For multi-turn agents, caching and routing reduce redundant computation while keeping memory costs bounded. On the training side, optimized parallelism and sequence packing speed up SFT and RL steps.

Billing is per GPU-minute for every model resource. According to Together, thanks to optimizations in its dedicated and multi-LoRA serving stacks, overall costs drop below token-based pricing as soon as experiments scale up. Several LoRA experiments can share one model resource, so a sweep does not require reserving more GPUs for each setup.

Deployment and getting started

Together has added multi-LoRA adapter serving in Dedicated Model Inference. Instead of starting a separate production deployment per run, users can swap adapters in a single GPU deployment to compare outputs and lower serving costs. The company also says each checkpoint retains its quality in serving, even after moving from multi-LoRA sampling to merged inference in FP4 precision.

To begin, developers install the latest Together SDK and read the docs. Together also offers cookbooks at github.com/togethercomputer/post-training-cookbook, covering GRPO training on the GSM8K benchmark, DPO as a custom loss, and multiple LoRA adapters within one model resource. For teams already using post-training services such as Tinker API, the company says it offers a simple migration path.

For teams weighing custom models, the pitch is that they can keep control of data, environments and objectives while outsourcing the infrastructure work. Whether the cost and quality claims hold will depend on individual workloads, as the source provides no benchmark figures.