Applied Compute Releases AC2, an AI Model Training Platform for In-House Research

Applied Compute has officially launched AC2 (Applied Compute Agent Cloud), the internal platform its research team used to build state-of-the-art models for clients such as Microsoft, NVIDIA, and Cognition. AC2 provides a single platform for AI teams to train open models on their own data, serve them at scale, and incorporate production data into the next run.

Building a Model Factory and Flexible Choice

AC2 enables customization of the models themselves, going beyond the prompt level. It supports the latest open models, allowing users to select the optimal base model based on performance, speed, and cost for each workload, and to switch quickly when new weights are released without rebuilding the stack. Additionally, training can be started with just a few dozen lines of code, allowing researchers to focus on experiments and data rather than infrastructure.

Engineered Experiment Process and Automation

Since RL model training is both a science and an art, the AC2 console acts as a cockpit where researchers can debug issues through rollout analysis, run comparisons, and grader iterations. The frontier-grade post-training stack is designed to maximize GPU performance while maintaining ML stability, adapting workloads as compute and context length scale. Furthermore, an applied research agent named Ari analyzes results and discovers failure modes throughout the training loop, monitors run health even when offline, reads rollouts, and takes actions to convert them into better data for the next experiment.

Deploying Dedicated Inference Capacity and Continuous Improvement

The serving stack balances latency, throughput, and cost, automatically scaling with traffic changes. Because training and serving occur on the same platform, production endpoints use the exact sampling configurations, numerical precision, and kernels used during training. Checkpoints from completed runs can be deployed in minutes with 99.9% uptime, auto-scaling, and low-latency inference based on speculative decoding. New checkpoints can also be deployed in minutes behind the same endpoint without changes to APIs, routing, or access controls.

The most valuable data, such as real user interactions, corrections, and feedback that occur after deployment, is collected and utilized for continuous training. Production traces identify failure modes and build new data points, while AC2's on-policy self-distillation generates learning signals from production traffic to pass to the next model, even if the original environment cannot be reproduced.