Together AI Launches Post-Training Service

Together AI on 19 October 2026 launched Together Post-Training, a service for building custom models on its platform through reinforcement learning from environment rewards (RL) and supervised fine-tuning (SFT). Unlike its existing fine-tuning product, the entire training loop is the customer’s to define, while Together runs the GPUs and infrastructure underneath. For companies trying to turn cheaper open-weight models into specialist agents, that removes much of the infrastructure work that stalls such projects.

The launch sits in a crowded space: post-training is how many AI-native firms now try to match frontier-model quality at lower serving cost. Together’s pitch is control plus correctness, with dedicated GPU allocations and a promise to keep training and inference numerically aligned.

In brief

Together AI launched Together Post-Training for custom RL and SFT on open models, starting with Muse Spark Open, GLM 5.3, Kimi K3 and Qwen 3.8-27B.
Billing is per GPU-minute, with an introductory rate of $9 per B200 GPU-hour for the first eight weeks; Tinker API users can migrate by changing one line.
Customer results quoted in the launch, from Deel, Jack & Jill and Collinear AI, are company-supplied testimonials, not independent evaluations.

Your loop, Together’s GPUs

Experiments run on Model Resources, GPU allocations reserved for one customer rather than shared with other users’ requests. Each resource can hold one or more Training Sessions on the same base model, with a Trainer for forward and backward passes and Generator instances for rollouts. SFT jobs can run with zero generators to save money, and asynchronous RL is supported to keep both sides busy during long agentic rollouts.

Together supports LoRA training, including several LoRA adapters sharing the same GPU capacity, plus full-weight training for select models. Built-in losses include cross-entropy, GRPO, CISPO and DPPO, and developers can compute a custom objective locally and send its gradients to the Trainer. Customers can bring any data, environments or agent harnesses and add their own steps, such as logging or external benchmarks.

The company says it prioritised output quality, targeting mismatches between training and inference engines that can make RL runs underperform or collapse. It also says checkpoints keep their quality when served in merged FP4 precision, and it has added multi-LoRA adapter serving to its Dedicated Model Inference product so teams can compare runs on one deployment.

Pricing and migration

Together bills each Model Resource per GPU-minute and argues that, at scale, training costs drop below token-based pricing. For the first eight weeks, the service costs $9 per GPU-hour of B200 usage. Developers already using Thinking Machines’ Tinker API can switch by replacing one import line, according to Together’s migration guide.

The launch model list is Muse Spark Open, GLM 5.3, Kimi K3 and Qwen 3.8-27B, and Together says more will follow in the coming weeks. Cookbooks on GitHub cover GRPO on GSM8K, DPO as a custom loss and training several LoRA adapters in one resource.

What early users say

Together quotes three early-access customers, all supplied by the company. Anthony Li, associate director of data science at Deel, said the service let his team run “up to five concurrent RL experiments on the same GPU capacity”. Oliver Normand, founding engineer at recruiting start-up Jack & Jill, said: “We matched a closed-source model’s quality in offline evaluations, using an open-source model that’s significantly cheaper to serve.”

Adit Jain, a researcher at Collinear AI, said his team had worked with Together since the Post-Training beta began in April. None of the three published evaluation numbers, and the Jack & Jill comparison is described only as an offline evaluation against an unnamed closed model.

Our read

Together is betting that custom post-training is becoming routine rather than a frontier-lab luxury, and the cheap introductory B200 rate plus one-line Tinker migration is a direct play for developers already doing it elsewhere. The open question is whether the quality promises hold up outside hand-picked early users. Watch for published before-and-after evaluations from customers — and what happens to demand when the $9 rate ends after eight weeks.

How we sourced this

Primary: Together AI blog (published 19 October 2026). Product details, pricing and model list are from that post; customer quotes are testimonials supplied by Together and have not been independently checked.