Post-training has become a continuous, central workload for agentic AI systems rather than a one-time finishing step, sh
Like elite athletes who continuously refine their skills between games, agentic AI models must keep adapting as environments shift, edge cases emerge and tools change. Unlike generative models that respond to prompts, agentic models must plan, use different tools and recover from problems they encounter mid-run.
Post-training, the phase that refines a model after initial training on raw data, is no longer a one-time finishing step. It's continuous, because the environment that agentic models operate in shifts fast. The tools an agent uses can change week to week, and edge cases surface in production that no test set anticipated. Each deployment brings its own codebase, policies and environment. Post-training runs loop back from production as new problems surface, creating a new compute pattern where the runs never stop. The compute footprint grows not because any single run is larger, but because the runs are continuous.
The goal of post-training is to maximize intelligence per dollar by maximizing the yield of every forward and backward pass in the continuous learning cycle. The forward pass, or inference, is measured in cost per token. Post-training is where intelligence is built. In pretraining, the model learns to predict the next token, giving it fluency but not intelligence. Post-training is where it learns to write code, plan multistep tasks, use a search tool and recover when something goes wrong. Because there's no answer key to memorize, only a reward, the model learns through reinforcement learning techniques. When given a task, it writes out an attempt, which is scored, and the lesson updates the model's weights across millions of attempts until intelligence grows.
Cost per token measures operating yield, while intelligence per dollar measures whether the investment in model intelligence is paying off. The two are nested, not competing. AI infrastructure that lowers cost per token also lowers the cost of every point of intelligence built into the model, and every point of intelligence built in raises the value of every token the inference factory serves.
NVIDIA Nemotron 3 Ultra, an open weight 550-billion-parameter mixture-of-experts model, offers verifiable benchmarks and a fully disclosed post-training recipe. It scored 71.7 percent on SWE-bench verified, a standard real-world coding benchmark where it produced a working fix for roughly seven in ten real software bugs from open source projects, each one checked against the project's own tests.
The NVIDIA Blackwell platform lowers cost per run and makes the frequent post-training the agentic era demands economically viable. The NVIDIA Vera Rubin platform extends this further, training the largest models with one-fourth the GPUs of the Blackwell generation. It was codesigned to maximize intelligence per dollar for the agentic post-training load with more rollouts per run, more environments in play and post-training cycles that never stop.
Prime Intellect continuously post-trains frontier open models on NVIDIA Blackwell and uses NVIDIA Dynamo for inference orchestration. With Vera Rubin, Prime Intellect plans to scale reinforcement learning environments and accelerate training-to-inference iteration loops. Prime Intellect has optimized its sandbox infrastructure to integrate with NVIDIA Vera CPUs, delivering on average 30 percent greater throughput per CPU compared to alternative x86 architectures.
Perplexity's reinforcement learning post-training stack runs asynchronously across hundreds of NVIDIA GPUs with an RDMA-based weight transfer engine that syncs trillion-parameter models in under two seconds between training and inference compute nodes. Together AI provides post-training as a service, including supervised fine-tuning, reinforcement learning and direct preference optimization on its AI Native Cloud platform, with plans to harness the Vera Rubin platform.