Build your model factory
Turn your data into better, faster, and cheaper models you own.
Choose your training path
Post-train your models yourself, with our researchers alongside you, or hand the whole loop to our team. Every path runs on the same Agent Cloud and ends with weights you own.
Managed
Your team drives the runs, environments, and graders on AC2. Every checkpoint deploys to dedicated inference in minutes, and the traffic it serves comes back as your next training set.
- For applied ML and product teams
- RL, SFT, preference optimization, and self distillation
- Bring your own harness, so training matches production
- Serve any checkpoint on dedicated inference
Embedded
Our researchers embed with your team and hand back a model that beats the frontier on your own benchmark. You keep the weights.
- For all teams, supported by our forward deployed researcher's
- Custom evals written on your product metrics
- Dedicated Blackwell capacity with SLAs, in our cloud or yours
- Full parameter training, and the weights are yours
Built and used every day by frontier researchers
AC2 is one unified platform to train, serve, and continuously improve models you own, with our researchers working beside your engineers.
The full stack for model customization
Launch with a single command
Select a model, dataset, environment, and grader. Start from any open model, and swap it as better ones ship.
Watch your model as it trains
Live monitoring, evals, and token level metrics for every run. Spot regressions automatically with our research agent, Ari.
Optimize toward your evals
Calibrate rewards to your specific product KPIs and domain expertise.
Online Training
Train with real production traffic, so your model gets better the more it's used.
Frontier post-training
Run experiments on less compute, iterate faster, and achieve higher performance
Bring your own harness.Your agent's learning environment should be faithful to production. Train inside the harness you already run.
Fully asynchronous RL.Our research on low-precision training recipes, high-throughput long-context inference, and fully async RL makes every run efficient and stable.
Full parameter fine tuning at frontier scale.Train multi-trillion parameter models with zero setup. Scale to thousands of GPUs without managing a cluster.
Continual learning with self-distillation.Methods like OPSD learn directly from non-replayable production traces so usage automatically becomes training signal.
Your AI research agent
Spend less time operating the research loop with Ari, applied research intelligence. Run iterative experiments automatically.
Ari watches the same metrics your researchers do and alerts you the moment a run drifts then fixes the config and relaunches.
Industry leading teams post-train with Applied Compute
News and Research
Enriching Domain-Specific Post-Training with Multi-Teacher On-Policy Distillation
Train your own decision model in AC2
Scaling On-Policy Self-Distillation with Hint Auto-Optimization
Start training today
State-of-the-art on your benchmarks, at a fraction of frontier cost, and yours to keep.