Applied Compute opened private beta access to AC2, a platform that lets AI teams train open-weight models on their own data, serve those checkpoints in production, and feed live traffic back into the next training run. The company says AC2 is the same internal system its researchers already use to build custom models for Microsoft, Nvidia, Cognition, Mercor, DoorDash and Harvey.
The pitch is that picking the best off-the-shelf model is no longer a durable edge, since open weights improve on a near-monthly cadence. Applied Compute is instead selling a repeatable pipeline: a company’s proprietary data becomes a continuous stream of model improvements rather than a one-time fine-tune. That is a bet on process over product, and it puts Applied Compute in competition with a growing list of vendors, including Together AI, Fireworks and Baseten, that already sell training-and-inference infrastructure for open models. Custom-model platforms are becoming a crowded category, and AC2’s differentiation rests on tying training and serving to the same underlying stack rather than stitching together separate tools.
Applied Compute says the platform works with the current crop of open-weight models and does not force teams onto a fixed training pipeline: engineers can wire in whatever harness they already use, and the company claims setup requires only a modest amount of custom code, not a rebuilt stack. It also points to a debugging console, an interface it says lets researchers step through individual rollouts and line up one training run against another, and a separate internal system, named Ari, that Applied Compute describes as watching jobs for failures and intervening on its own when no engineer is present. None of these claims come with independent benchmarks or published performance numbers; the announcement is Applied Compute describing its own product.
On the serving side, Applied Compute says a finished checkpoint can go live in production in a short window, still reachable at its original endpoint, backed by autoscaling and a target the company puts at 99.9 percent uptime. The company also says the version answering live traffic runs with the identical numeric settings used during training, which it frames as a way to close the gap between how a model scores in evaluation and how it behaves once real users hit it, a gap that has tripped up other teams shipping fine-tuned models. Applied Compute has not disclosed pricing, compute costs or which open-weight model families are supported at launch.
The more novel claim is what Applied Compute calls on-policy self-distillation: capturing production traces and user corrections and turning them into training signal for the next model run, even in cases where the original setup that produced a rollout no longer exists to run it again. If it works as described, that would let a customer’s live traffic function as a data flywheel without requiring a fully reproducible simulation. Applied Compute has not published data showing how much this improves model quality over standard fine-tuning, and the claim should be treated as unverified until customers or independent researchers report results.
AC2 is available only in private beta, and Applied Compute is directing interested teams to request a demo rather than self-serve. For engineering leaders evaluating whether to build in-house fine-tuning infrastructure or buy it, AC2 adds one more vendor to a list that already includes hyperscaler-native options from AWS, Google Cloud and Azure, meaning the decision now hinges less on raw capability and more on how tightly a vendor’s training loop ties back to a company’s own production data.
Applied Compute announced AC2 in a product page published August 25, 2026.
