Abundant runs more than two million AI rollouts on Modal each month

Abundant has built its reinforcement learning rollout platform primarily on Modal, running more than two million Sandboxes per month as of November 2026. The company says the setup helps it create and evaluate difficult tasks for AI labs without managing EC2 clusters directly.
- Abundant has created over 75,000 reinforcement learning rollouts.
- The company runs over two million Modal Sandboxes per month as of November 2026.
- Its rollout monitor sends a heartbeat every 30 seconds.
- Modal bursting lets Abundant accommodate resource peaks with lower requests.
- Frontier RSI used over 1,300 hours across nearly 500 agent trials.
Infrastructure for reinforcement learning tasks
Abundant builds reinforcement learning environments used by frontier labs including OpenAI, Anthropic, Google, and Meta. Its tasks are designed to test agent capability: a model receives instructions and a goal, attempts the task, and then a verifier grades the result as pass or fail.
Each task also requires infrastructure to provide compute, Dockerfiles, isolation, and, in some cases, accelerators such as GPUs. Abundant has created over 75,000 rollouts, ranging from GPU kernel optimization to long-horizon coding, with each rollout running in a Modal Sandbox.
Modal Functions power Abundant’s scheduler API, which starts tasks, cleans up dead runs, and reviews rollouts. Modal Sandboxes provide the rollout environments. According to Modal, Abundant runs over two million Sandboxes per month as of November 2026.
A monitor inside each Sandbox
Abundant initially evaluated Sandbox providers because it wanted to avoid managing EC2 clusters directly. The company says Modal allowed it to build infrastructure in code, launch Sandboxes through an API, and scale traffic as demand changed.
Each rollout launches a task, lets an agent attempt it, verifies the work, and reports the result. A monitor must also check that the Sandbox remains active. Abundant chose to place that monitor inside the Sandbox rather than host it separately, while using short-lived Modal Functions to start Sandboxes.
The monitor sends a heartbeat every 30 seconds, along with live events and the final result, through HTTP calls. If a Sandbox becomes quiet, the control plane attempts to connect three times before terminating it.
“Modal lets us focus on our core business without worrying about whether our infrastructure will work. We’re able to work in code and scale our traffic based on customer demands in a given moment,” said Kyle van den Bergh, a Member of Technical Staff at Abundant.
Bursting reduces the need to provision for peak usage
Abundant’s tasks did not consume CPU and memory at consistent rates, creating a challenge for resource allocation. Modal recommended using its bursting feature, which allows Sandboxes to automatically size up and back down when they need additional resources, if capacity is available on the machine.
With bursting, Abundant could request resources based on average usage while still accommodating peaks. Modal says this can help Sandboxes avoid CPU throttling and out of memory errors. Lower resource requests also reduced the idle compute Abundant was paying for and made it easier to support a wide range of tasks without choosing CPU and memory allocations for each Sandbox individually.
Verifier design protects benchmark results
Abundant also focused on making its verifiers resistant to reward hacking, where a model finds a way to achieve a favorable result without completing the intended task. During early testing, the company found that models sometimes located copies of the task instructions, creating evaluation awareness and skewing results.
To address this, Abundant changed how runs start so that instructions are deleted before work is completed. It also added a reviewer agent to the verification step. That agent reviews traces and flags signs that a model knows it is being evaluated.
Modal says the rollout architecture took around two weeks to implement. The resulting infrastructure supports Abundant’s focus on task verification and on creating difficult, open-ended evaluations for AI labs. Its Frontier RSI benchmark, which tests an agent’s ability to improve itself, used more than 1,300 hours of agent research time across nearly 500 agent trials to produce its initial published results.


