Together ties PyTorch RL to cloud

This partnership shows Together is trying to own the full model lifecycle, not just cheap inference. Reinforcement learning is the stage where teams run thousands of model rollouts, score outputs, and push weight updates back into training. By helping build an open source RL framework with the PyTorch team at Meta and wiring it into its cloud, Together makes its GPUs and training stack the default place to run that loop.1 2

• The announced framework is meant for agentic AI, which means training models by letting them use tools, browse, write code, or complete multi step tasks, then learning from success and failure. That moves Together from hosting models after they are built into the workflow where model behavior is actually improved.1 2
• The strategic value is distribution through the PyTorch ecosystem. PyTorch is one of the core open source layers for AI training, and Meta has continued building public agentic and RL tooling around it. If the RL framework feels native to PyTorch, developers can plug Together into existing training code instead of adopting a separate closed stack.1 4 5
• This also puts Together in a race with newer RL infrastructure plays. OpenPipe has positioned serverless RL as part of a vertically integrated stack with CoreWeave and Weights & Biases, while Scale AI has launched RL environments for enterprise agent training. Together is differentiating by tying open source software directly to its own cloud capacity and training APIs.6 7

The next step is for RL to become a normal cloud workload, like fine tuning already is. If Together keeps turning open source training components into managed platform products, it can capture more spend from labs and startups as they move from serving models to continuously training agents that get better in production.1 6