Applied Compute

Applied Compute is launching AC2, its Agent Cloud for training, serving, and continuously improving open models. The platform is aimed at enterprise AI teams that want to post-train their own frontier agents on top of open weights like Qwen3 rather than depending on closed model APIs. Teams register a dataset, wrap their own agent harness in a lifecycle interface (the demo uses OpenCode for a code Q&A agent), define a reward function, and launch a run from the CLI, with preset configs for popular open models and post-training algorithms including GRPO, SFT, and OPSD. Trained models can then be deployed to an autoscaling inference endpoint, and traces captured in production feed back into further training through techniques like self-distillation. A notable piece of the launch is Ari, an in-house research agent that watches runs, reads rollouts, and takes actions such as restarting a job when a remote tool server fails and reward collapses, with an inspectable execution trace and cross-run reporting. Applied Compute is led by co-founder and CEO Yash Patil, previously at OpenAI and Stanford , alongside co-founders Rhythm Garg and Linden Li. Yash was a key member of the Codex agentic software engineering effort, Rhythm contributed to the first RL-trained reasoning model (o1), and Linden specialized in ML systems and infrastructure for RL training . This launch matters because the market is splitting between teams that rent frontier intelligence from a handful of labs and teams that want to own a model tuned to their data, tools, and workflows. Applied Compute is positioning AC2 as the substrate for the second group, and it says the system already powers custom training work with Microsoft, NVIDIA, Cognition, Harvey, Mercor, and DoorDash. For founders and operators evaluating whether to build in-house post-training capacity, AC2 is worth a look as an integrated path from dataset to RL run to served agent without stitching together separate GPU, orchestration, and eval stacks.