Together AI Launches Post-Training Service for Custom AI Models

Together AI has launched Together Post-Training, a service that lets developers build custom AI models with their own reinforcement learning and supervised fine-tuning loops while the company runs the underlying GPU infrastructure. The launch, announced on October 19, 2026, targets a bottleneck that has quietly become one of the defining constraints of the agent era: Post-Training RL infrastructure is getting more complex, and many teams that could benefit from it either never start or stall under the weight of the plumbing, according to the company's announcement.

The pitch is simple to state and hard to execute. Custom model behavior — especially for agents that plan, use tools, and recover from mistakes over long horizons — is increasingly produced in Post-Training, not in pretraining. But running RL at any serious scale means standing up rollout infrastructure, keeping training and inference engines numerically consistent, managing GPU capacity across trainers and generators, and iterating on environments and reward signals. Together AI is productizing that whole layer as a managed service, with introductory pricing of $9 per GPU-hour of B200 usage for the first eight weeks, according to Together AI's official announcement.

Post-Training Puts the Training Loop in the User's Hands

The service is organized around what Together AI calls Model Resources: GPU allocations dedicated to a single customer, with no contention from other users' workloads. Each Model Resource can hold one or several Training Sessions, where each session is a single experiment on the same base model — for example, one set of LoRA adapters. Because the capacity is dedicated, teams can also run full-weight fine-tuning beyond LoRA when adapting models to large knowledge-intensive tasks, reported Unite.AI in its detailed breakdown of the launch.

Each session pairs a Trainer instance, which handles forward and backward passes and weight updates, with Generator instances that execute rollouts from the policy. Supervised fine-tuning runs can start with zero generators to save cost, and users control when trainer parameters are pushed to the generators. Asynchronous RL is supported for high utilization of both components, which matters most during long agentic rollouts where generation and optimization would otherwise take turns idling. The service supports LoRA training — including multiple adapters sharing the same GPU capacity — and full-weight experiments for select models, with standard losses such as cross-entropy, GRPO, CISPO, and DPPO. Teams can also compute a custom training objective locally and send its gradients to the Trainer, according to Unite.AI.

The SDK surface is deliberately granular. Developers provision a Model Resource — the billed unit — attach a training session, then drive training with per-step calls: sample_batch to generate completions for scoring, forward_backward to compute the loss and backpropagate, optim_step to update weights, and weights_sync to publish new weights to generator replicas. SFT and DPO skip the sampling and sync steps since they do not generate completions. Billing is per GPU-minute rather than per token, which Together AI says makes training cheaper than token-based pricing once experiments scale. Crucially, runs stay under the user's control mid-flight: per-step metrics can be inspected, the learning rate changed, data reshuffled, or the run stopped on custom logic — a contrast with job-style fine-tuning services that proceed to completion without intervention.

Onboarding is aimed at working teams rather than greenfield experiments. Together AI published a post-training-cookbook repository on GitHub with worked examples covering GRPO training on the GSM8K benchmark, DPO implemented as a custom loss, and multiple LoRA adapters trained within one Model Resource. Teams with an existing setup on the Tinker API get a migration path that replaces a single import line, lowering the switching cost for the shops most likely to adopt early.

Why Agentic Teams Are the Target Audience

The framing of the launch makes clear who Together AI is courting: teams building agents. Customers define their own data, environments, and agentic harnesses, and can add steps such as logging or external benchmarking while Together AI co-optimizes the underlying systems for model quality and training speed. Rollout engines draw on the same optimizations as the company's serving stack, with intelligent caching and routing for multi-turn agents, and parallelism and sequence-packing techniques that accelerate both SFT and RL steps.

Training stability gets explicit attention, and for good reason. The company says the service eliminates common sources of policy drift between training and inference engines, where tiny numeric mismatches from identical weights can cause an RL run to underperform or collapse — using techniques such as Rollout Routing Replay alongside its training-stability research. Anyone who has watched an agentic RL run diverge late in training will recognize the problem being named here; drift between the engine that rolls out and the engine that learns is one of the quiet killers of long-horizon agent training.

Early-access users illustrate the intended workload. Anthony Li, associate director of data science at Deel, said his team began running real RL experiments within a couple of hours and ran up to five concurrent RL experiments on the same GPU capacity. Oliver Normand, founding engineer at Jack & Jill, said his team post-trained a model predicting which candidates a hiring manager will want to meet — a core step in the search performed by the company's Jill recruiting agent — and matched a closed-source model's quality in offline evaluations using an open-source model that is significantly cheaper to serve. Adit Jain, a researcher at Collinear AI, described the platform as intuitive, flexible, and powerful for scaling sophisticated agentic workloads, and said seamless self-onboarding made it easy to launch full-scale agentic RL runs. The testimonials were shared by Together AI in its launch materials and have not been independently verified, as Blockchain.News noted in its coverage.

From Checkpoint to Production in FP4

The deployment path is part of the product, not an afterthought. Together AI now supports multi-LoRA adapter serving in its Dedicated Model Inference product, letting users swap adapters within a single GPU deployment instead of spinning up a separate production deployment per experiment. Each training checkpoint is said to retain its quality in serving even after switching from multi-LoRA sampling to merged inference in FP4 precision, with a preparation step required for GLM-5.3 and Kimi-K3 adapters. The documented workflow is train, evaluate, productionize: train several adapters against one frozen base model, attach the candidates to a single multi-LoRA deployment and compare them by naming a different adapter per request, then ship the winner in merged mode on its own deployment.

The launch lineup is entirely open-weight: Muse Spark Open, GLM 5.3, Kimi K3, and Qwen 3.8-27B, with Together AI saying the selection will expand in the coming weeks to cover both frontier LLMs and efficient smaller models. For companies that want a partner rather than a platform, forward-deployed engineers work alongside customer teams across the full lifecycle, from selecting an open model and agent harness through data and training environments to evaluations and shipping an optimized endpoint.

The timing fits a company converting capital into capacity. Together AI said in July 2026 that it had raised an $800 million Series C and secured commitments for more than 500 MW of independently capitalized compute capacity, reported Blockchain.News — the kind of balance sheet that makes dedicated-GPU product lines credible. The broader market context is a compute race: fresh AI chip financing is expanding the hardware options for exactly this kind of workload, while cloud platforms race to serve enterprise agentic AI infrastructure. If Post-Training is where agent capability is actually manufactured, the companies that make it boring and reliable own a strategic layer. Together AI's bet is that the next wave of model differentiation happens after pretraining — and that teams would rather write training loops than babysit GPU clusters.